JP2011181021A - Time-series signal identification device, time-series signal identification method and program - Google Patents

Abstract

A time-series signal identification device, a time-series signal identification method, and a program capable of efficiently and accurately identifying time-series signals composed of three or more classes are provided.
A discriminator generation device 1 captures learning data A in which a correct class is associated with a time-series signal. Then, an identification score r _{t, k} indicating the degree of coincidence between the result of identification of the learning data A by the weak classifier h _{t, k} and the correct class of the learning data A is calculated, and the weakness is obtained so that the identification ability is maximized. The discriminator 100 is generated by setting the discriminator _{ht, k} . The time-series signal identification device 2 obtains the identifier 100, and generates a strong classifier H _k for each class identifies the class of the identification target data (time series signal).
[Selection] Figure 1

Description

  The present invention relates to a time-series signal identification device, a time-series signal identification method, and a program for classifying time-series signals such as moving images and sounds.

  Techniques for identifying which of a plurality of classes a time-series signal such as a moving image or sound has been proposed as follows. Here, for example, if the class is a video signal, “the driver of the car is reaching out to the rearview mirror” “the driver of the car is reaching out to insert the engine key” In the case of an audio signal, it represents the type of an individual word such as “A” or “I” or a word having a meaning such as “search” or “re-search”.

  Common identification algorithms that do not rely on time-series signals include Support Vector Machine (SVM) (see Non-Patent Document 1), Adaboost (see Non-Patent Documents 2 and 3), and the like. A hidden Markov model is known as an identification algorithm for time series signals (see Non-Patent Document 4). When executing these techniques, a normal recognition method is to input a single time-series signal of interest from the beginning to the end and estimate an optimum recognition result from information of the entire signal.

  On the other hand, in many situations where the class of a time-series signal is actually identified and the identification information is used, the time-series signal is input before the end of the series of time-series signals, that is, until the end. It is necessary to recognize the class to which the time-series signal belongs from the initial signal as much as possible without waiting for it. For example, from the time-series signal being input, identify the class that "the driver of the car is reaching out to insert the engine key", and use the identification information to light the keyhole to start the engine It can be applied to a scene where a human motion is recognized and assisted. The problem of recognizing this time series signal class as early as possible is called the early recognition problem.

  For this early recognition problem, an Earlyboost (early recognition boosting) method (see Non-Patent Document 5) has been proposed. In the Earlyboost method, a discriminator satisfying as high a discriminating performance as possible is selected from only observation information in time frame 1. Subsequently, using the information of time frame 2, the discriminator is updated so that the time series signal that has failed to be identified by the information of time frame 1 can be identified as correctly as possible. In this way, it is possible for each time frame of the time-series signal by updating the discriminator that failed to be identified in the time frame immediately before the time frame to be identified so that it can be correctly identified. The classifier with the maximum recognition performance is constructed.

  In the identification method described above, the number of classes to be identified is limited to two. That is, it is limited to the identification of whether or not it belongs to a certain class. In this regard, Adaboost.MH (Non-Patent Document) is an identification method in which Adaboost (see Non-Patent Documents 2 and 3), which is the above-described non-time-series signal identification method, is extended to three or more classes (multi-class) identification. 6) has been proposed.

Burges, C. J. C., “A Tutorial on Support Vector Machines for Pattern Recognition”, Data Mining and Knowledge Discovery, Vol. 2, No. 2, pp. 121-167, 1998. Freund, Y. and Schapire, R. E., “A Decision-theoretic Generalizing of On-line Learning and an Application to Boosting”, Journal of Computing Systems and Science, Vol. 55, No. 1, pp. 119-139, 1997. Yasushi Yagi (Editor), Hideo Saito (Editor), Computer Vision Advanced Guide <1> Level Set, Graph Cut, Particle Filter, Tensor, AdaBoost (CVIM Tutorial Series), Chapter 5, Adcom Media (2009) L. R. Rabiner, “A Tutorial on Hidden Markov Models and Selected applications in Speech Recognition”, Proceedings of the IEEE, Vol. 77, No. 2, pp. 257-286, 1989. Uchida, S. and Amamoto, K., “Early Recognition of Sequential Patterns by Classifier Combination”, in Proceedings of the International Conference on Pattern Recognition (ICPR), 2008. Schapire, R. and Singer, Y., “Improved Boosting Algorithms using Confidence-rated Predictions”, Machine Learning, Vol. 37, No. 3, pp. 297-336, 1999.

  However, since the technique based on SVM, Adaboost, and Adaboost.MH described above is an identification method related to a non-time series signal, it cannot correspond to identification of a time series signal. The technique based on the hidden Markov model is based on the assumption that a single time-series signal is input from the beginning to the end and the optimal recognition result is estimated from the information of the entire signal. Early recognition of the class to which the signal belongs cannot be performed. The Earlyboost method can only be applied to early recognition of two classes. In other words, when a certain time series signal is identified, it can only determine whether or not it belongs to a certain class. Therefore, it cannot be used as it is for an actual identification problem that requires more complicated identification between multiple classes.

On the other hand, when a two-class identification model such as the Earlyboost method capable of early recognition is applied to multiple classes of three or more classes, the following method can be considered. (1) When 3 or more K classes are set, for each of the K classes, a classifier that solves the 2-class identification problem of “class of interest” and “other class” is prepared. And compare their discrimination scores. (2) A two-class classifier is configured for all class pairs using the binomial coefficient _k C ₂ , two-class classification is performed by a plurality of classifiers, and the target time-series signal is recognized by combining the results.

However, these methods have the following problems. First, in the method (1), since a plurality of classes that have inherently different distributions are aggregated into one class called “other classes”, the discrimination function becomes complicated, and it is difficult to obtain an accurate discriminator. . On the other hand, in the method (2), such a problem does not occur, but a large amount of classifier learning and recognition processing is required by the pattern of _k C ₂ , so that the calculation amount is enormous.

  The present invention has been made in view of such a background, and the present invention is capable of efficiently and accurately identifying a time-series signal composed of three or more classes, a time-series signal identification device, and a time-series signal identification. It is an object to provide a method and a program.

  In order to solve the above-mentioned problem, the invention according to claim 1 is characterized in that, for each time frame obtained by dividing the time series signal every predetermined time, three or more time series signals that are sequences in which identification target data in the time frame are arranged are arranged. A time-series signal identification device for classifying a class of data, a first input unit that receives an input of a sequence of identification target data from the beginning of the time frame to a predetermined time frame shorter than the time-series signal; A strong classifier generating unit that generates, for each class, a strong classifier that calculates the degree to which the series of identification target data received by the first input unit belongs to the class, and the class generated by the strong classifier generating unit Using each strong classifier, the degree to which the series of identification target data received by the first input unit belongs to the class is calculated, and belongs to the class calculated for each class A classifier of a classifier to which the classifier of the strong classifier showing the maximum value of the match belongs to the class to which the time series signal belongs, and to which the time series signal belongs without identifying all the time series signals A time-series signal identification device characterized in that

  According to the sixth aspect of the present invention, for each time frame obtained by dividing the time series signal every predetermined time, a time series signal that is a series in which identification target data in the time frame is arranged is identified into three or more classes. A time-series signal identification method used in a time-series signal identification apparatus, wherein the time-series signal identification apparatus is configured to identify a sequence of identification target data from the beginning of the time frame to a predetermined time frame shorter than the time-series signal. Using the step of receiving an input, the step of generating for each class a strong classifier that calculates the degree to which the received series of identification target data belongs to the class, and using the generated strong classifier for each class, The degree to which the received identification target data series belongs to the class is calculated, and the maximum value of the degrees belonging to the class calculated for each class is calculated. The step of setting the class of the strong classifier to the class to which the time series signal belongs is performed, and the class to which the time series signal belongs is identified without identifying all the time series signals, A time-series signal identification method is used.

By doing so, the time-series signal identification device according to the present invention receives a sequence of identification target data from the beginning of the time frame to a predetermined time frame shorter than the time-series signal, and receives the sequence of identification target data received A strong discriminator that calculates the degree of belonging to a class is generated. Then, the strong classifier for each class calculates the degree of coincidence of whether or not the input identification target data belongs to that class, and the class of the strong classifier indicating the maximum value among the calculation results is input. It can be the class to which the identification target data belongs.
Therefore, according to the present invention, the time-series signal identification device can recognize three or more classes efficiently and accurately without identifying all of the time-series signals.

  The invention according to claim 2 is a weak classifier for identifying whether or not identification target data prepared for each time frame and each class belongs to its own class in the first input unit. , An identification weight representing the reliability of the weak classifier in each time frame is input, and the strong classifier generator generates the time series from the beginning of the time frame received by the first input unit for each class. The weak classifier for each time frame up to a predetermined time frame shorter than a signal is the strong classifier obtained by adding the weak classifier's identification weights. A time-series signal identification device was used.

  The invention according to claim 7 is a weak discriminator for identifying whether or not identification target data prepared for each time frame and each class belongs to its own class, and weak in each time frame. In the step of receiving an input of an identification weight representing the reliability of the discriminator, and generating the strong discriminator for each class, the time series signal from the head of the received time frame for each class The step of using the weak discriminator for each time frame up to a predetermined time frame shorter than the sum of the weak discriminators by the discrimination weight of the weak discriminator as the strong discriminator is further executed. The time-series signal identification method according to item 6 is adopted.

By doing in this way, the time series signal discriminating apparatus according to the present invention uses the weak classifier and the discriminating weight that are input, for each class, for each time frame up to a predetermined time frame shorter than the time series signal. A weak discriminator can be added to the discriminating weight of the weak discriminator to make a strong discriminator.
Therefore, according to the present invention, since the strong classifier considers the classification weight of the weak classifier in each time frame and identifies the class of identification target data, the time-series signal identification device has three or more classes. It is possible to identify the time-series signal with further improved accuracy.

  According to a third aspect of the present invention, for each time frame, learning in which a time series signal as identification target data for learning of the time frame is associated with a correct class to which the identification target data for learning belongs is associated. A second input unit that accepts input of data, and a time-series signal of each time frame in the learning data accepted by the second input unit is acquired in order of oldest time frame, and the time series of the acquired time frame For each of the classes, the identification in the time frame is such that the value of the signal increases as the identification result of the weak classifier matches the correct class to which the time-series signal of the time frame indicated in the learning data belongs. An identification score calculation unit for calculating a score, and for each time frame and for increasing the calculated identification score A weak classifier learning unit that sets the weak classifier for each class, and an identification that sets an identification weight for each time frame so that the value increases as the identification score for each time frame and for each class increases. A weight setting unit, and the weak classifier configured to input the weak classifier set by the weak classifier learning unit and the classification weight set by the identification weight setting unit to the first input unit. The time series signal identification device according to claim 2, wherein the time series signal identification device is the identification weight.

  In the invention according to claim 8, for each time frame, a time series signal as identification object data for learning of the time frame is associated with a correct class to which the identification object data for learning belongs. Receiving the learning data input, and acquiring the time-series signal of each time frame in the received learning data in the oldest order of the time frame, and for the time-series signal of the acquired time frame, the weak classifier Calculating an identification score in the time frame for each of the classes so that the value increases as the identification result matches the correct class to which the time-series signal of the time frame indicated in the learning data belongs; For each time frame and each class, the weak knowledge is increased so as to increase the calculated identification score. And further, the step of setting an identification weight for each time frame so that the value increases as the identification score for each time frame and for each class increases. 8. The time-series signal identification method according to claim 7, wherein the weak classifier and the set identification weight are the weak classifier and the identification weight that are input in the step of receiving input of the learning data. It was.

By doing in this way, the discriminator generating device according to the present invention uses the learning data in which the correct class is associated with the time-series signal, and the learning data by the weak discriminator set for each time frame and each class. It is possible to calculate the identification score so that the value becomes higher as the identification result of this matches the correct class of the learning data, and the weak classifier can be set to increase the calculated identification score. The identification weight can be set so that the higher the identification score, the larger the value.
Therefore, according to the present invention, the weight of the weak classifier can be considered for each time frame when identifying the class of the identification target data, and the weak classifier is optimally integrated from the viewpoint of the classification capability. A strong classifier can be generated.

  The invention according to claim 4 is the result of identifying the time series signal acquired in the oldest order of the time frame by the weak classifier set by the weak classifier learning unit, and the time frame indicated in the learning data. Are compared with the correct class to which the time-series signal belongs, and the error decreases so that the value decreases if the weak classifier identification result is correct, and the value increases if the weak classifier identification result is incorrect. An error weight calculator configured to calculate a weight, and to set the calculated error weight as the error weight for the learning data of the time frame next to the time frame in the acquired time series signal, the identification score calculator And the calculation is performed such that the value of the identification score in the time frame of the learning data having a large error weight calculated by the error weight calculator is large. And time-series signal identifying apparatus according to claim 3.

  The invention according to claim 9 is the result of identifying the time-series signal acquired in the oldest order of the time frame by the set weak classifier and the time-series signal of the time frame indicated in the learning data. The error weight is calculated so that the value decreases if the weak classifier identification result is correct, and the value increases if the weak classifier identification result is incorrect. And further executing the step of setting the calculated error weight as the error weight for the learning data of the time frame next to the time frame in the acquired time series signal, and calculating the identification score, 9. The calculation is performed such that the value of the identification score in a time frame of learning data having a large error weight is increased. And time-series signal identifying method according.

  Therefore, according to the present invention, the recognition result by the weak classifier is compared with the correct class of the learning data, and the error weight is set to be small if the recognition result is correct, and the error weight is set if the recognition result is incorrect. Set to be larger. By calculating the identification score in the next time frame using the error weight set in this way, the weak classifier that has mistakenly identified the class to which the learning data belongs is updated so that it can be recognized more correctly. And an identification weight can be calculated.

  In the invention according to claim 5, as a result of the identification score calculation unit identifying the time series signal of the acquired time frame by the weak classifier, time series data not belonging to the class does not correctly belong to the class. 5. The time-series signal identification device according to claim 4, wherein the identification score is calculated so that the identification score in the class and the time frame is high when the identification score can be identified.

  Further, in the invention according to claim 10, as a result of identifying the time series signal of the acquired time frame by the weak classifier in the step of calculating the identification score, the time series data not belonging to the class is correctly The time series signal identification method according to claim 9, wherein the identification score is calculated so that the identification score in the class and the time frame becomes high when it can be identified as not belonging to the class. .

  Therefore, according to the present invention, when the weak classifier can identify that the time-series data that does not belong to the class does not belong correctly, the weak classifier can be set with the identification score being increased. .

  The invention according to claim 11 is a program for causing a computer to execute the time-series signal identification method according to any one of claims 6 to 10.

  According to such a program, the time-series signal identification method according to any one of claims 6 to 10 can be executed by a general computer.

  ADVANTAGE OF THE INVENTION According to this invention, the time series signal identification device, the time series signal identification method, and program which can identify the time series signal which consists of three or more classes efficiently and accurately can be provided.

It is a functional block diagram which shows the structural example of the discriminator production | generation apparatus which concerns on this embodiment. It is a flowchart which shows the flow of a process of the discriminator production | generation apparatus which concerns on this embodiment. It is a functional block diagram which shows the structural example of the time series signal identification device which concerns on this embodiment. It is a flowchart which shows the flow of a process of the time series signal identification device which concerns on this embodiment. It is the average discrimination rate of the time series signal in the example of experiment using the time series signal discriminating device concerning this embodiment. It is the figure which compared the experimental result of the time series signal identification device which concerns on this embodiment, and the experimental result which applied Earlyboost (early recognition boosting) of a prior art.

  Next, modes for carrying out the present invention (referred to as “embodiments”) will be described in detail with reference to the drawings as appropriate.

≪Overview of operation of discriminator generator and time-series signal discriminator≫
First, the operation | movement outline | summary of the discriminator production | generation apparatus 1 which concerns on this embodiment, and the time series signal identification device 2 is demonstrated.

The discriminator generation device 1 according to the present embodiment takes in learning data A in which a correct class is associated with a time-series signal, and generates a discriminator 100 that stores weak discriminators h _{t, k} and discrimination weights α _t. It is a device to do. The time-series signal identification device 2 according to the present embodiment takes in the information of the classifier 100 and identifies the class for the identification target data (time-series signal) x. In addition, in this embodiment, although the function with which the discriminator production | generation apparatus 1 and the time series signal discrimination apparatus 2 are provided is demonstrated separately, the discriminator production | generation apparatus 1 is added to the time series signal discrimination apparatus 2 as one housing | casing. It can also be configured to include the provided functions.

The discriminator generation device 1 takes in the learning data A in the oldest order of the time frame, and the discrimination result of the learning data A by the weak discriminator _{ht, k} set for each time frame and each class and the correctness of the learning data A An identification score r _{t, k} indicating the degree of coincidence with the class is calculated. The weak classifier h _{t, k} is a function for identifying whether or not the time-series signal belongs to its own class. Then, the discriminator generation device 1 uses the calculated discrimination score rt _{, k} to weaken the discriminator ht _{, k} so that the discrimination score rt _{, k} is maximized, that is, the discrimination capability is maximized _. Adjust the _k parameter. Next, the discriminator generation device 1 determines the discrimination weight α _t indicating the reliability of the weak discriminator h _{t, k} in the time frame in consideration of the discrimination scores r _{t, k} of all classes.

Subsequently, when the classifier generating device 1 processes the next time frame, the weak classifier _{ht, k} in each class is small when the class of the learning data A is correctly recognized. The error weight D _{t, i, k} to be increased is set to the identification score r _{t, k} in the next time frame, and the identification score r _{t, k} is calculated. By repeating such processing, the weak classifier h _{t, k} and the identification weight α _t are set so as to correct the recognition error of the weak classifier h _{t, k in} the previous time frame. .

Then, the time-series signal discriminating apparatus 2 acquires the discriminator 100 (the weak discriminator h _{t, k} and the discriminant weight α _t ) generated by the discriminator generating apparatus 1 and generates the strong discriminator H _k for each class. . This strong classifier H _k calculates the degree of coincidence of whether or not the input time-series signal (identification target data) x belongs to its own class in consideration of the identification weight α _t in each time frame. Then, the class of the strong classifier H _k indicating the maximum value among the calculation results is identified as the class to which the input identification target data x belongs.

≪Classifier generator≫
Next, the discriminator generation device 1 will be specifically described. FIG. 1 is a functional block diagram illustrating a configuration example of a discriminator generation device 1 according to the present embodiment. The discriminator generation device 1 uses the input learning data A, constant information 131 and weak discriminator information (h _1,1 , h _1,2 ,..., H _{T, K} ) stored in the discriminator 100. 132 and identification weight information (α ₁ , α ₂ ,..., Α _T ) 133, as shown in FIG. 1, a control unit 10, an input unit (second input unit) 11, An output unit 12 and a storage unit 13 are included.

First, various data used in the discriminator generation device 1 will be described. In the present embodiment, R represents a set of one-dimensional real numbers, and R ^d represents a set of d-dimensional real vector. Also, t represents a time frame index (t = 1, 2,..., T).

The learning data A input to the discriminator generation device 1 includes a set of a time series signal x _i and a class variable y _i to which the time series signal x _i belongs. This class variable y _i is a correct class of the time-series signal x _i given in advance by the administrator of the discriminator generation device 1 or the like. Note that i is an index of the learning data A, and i = 1, 2,. Therefore, learning data A = {x _i , y _i } ^N _{i = 1} .

The time series signal x _i is a sequence of d-dimensional real vector x _{i, t} . The sequence length, that is, the length of the time series signal x _i is T in all i. That is, if the time frame index is t = 1, 2,..., T, x _i = {x _{i, t} εR ^d }. The maximum number of classes of the class variable y _i is K, and the value is given in advance. Therefore, the class variable y _i ε {1, 2,..., K}. Here, K numbers such as 1,..., K are used as symbols (class labels) representing classes. However, as long as K types of classes can be identified, symbol strings or the like may be used as class labels. Good. The learning sequence signal x _i when the data A can utilize the series signal x _i when that has been acquired in advance from the video camera and a microphone, or the like.

The discriminator generation device 1 uses the learning data A input via the input unit 11, and uses each constant relating to the learning data A (the total number N of learning data, the total number K of classes to be identified, the total number T of time frames, (1) Weak classifier h _{t, k} and (2) Discrimination weight α _t that are optimally set as information of discriminator 100 after storing dimension number d) of time series signal x _i in storage unit 13 I will remember it.

Here, when the weak discriminator _{ht, k} receives the value x _{i, t} in the time frame t of the time-series signal x _i and recognizes that the value belongs to the class k, the weak classifier _{ht, k} sets “1” to the class When recognizing that it belongs to other than k, this function returns “−1”. Further, the identification weight α _t εR, t = 1, 2,..., T is a weight for the weak classifier h _{t, 1: k} . Note that the time-series signal identification device 2 according to the present embodiment optimally sets the identification weight α _t of the weak classifier _{ht, 1: k} for each time frame t when identifying the class of the data to be identified. By doing so, it is possible to generate a strong classifier H _k that is optimally integrated with the weak classifiers _{ht, 1: k} from the viewpoint of discrimination capability.

Further, error weights D _{t, i, k} ∈R, t = 1, 2,..., T, i = 1, 2,..., N, k in the calculation process of the weak classifiers h _{t, k} and the identification weight α _t. = 1, 2,..., K, and identification scores r _{t, k} εR, t = 1, 2,..., T, k = 1, 2,.

(Device configuration)
Hereinafter, the configuration of the discriminator generation device 1 will be specifically described.
The input unit (second input unit) 11 includes an input interface for inputting information from the outside, receives input of learning data A and the like, and passes it to the control unit 10. The output unit 12 includes an output interface that outputs information generated by the discriminator generation device 1, and outputs various information stored in the discriminator 100 to the time-series signal identification device 2 and the like. In addition, the storage unit 13 stores various data (constant information 131, weak classifier information 132, identification weight information 133, error weight information 134, and identification score information 135) necessary for the processing of the control unit 10. As the parameters set optimally, constant information (N, K, T, d) 131, weak classifier information (h _1,1 , h _1,2 ,..., H _{T, K} ) 132, and identification weight information The classifier 100 in which (α ₁ , α ₂ ,..., Α _T ) 133 is stored is stored.

  The control unit 10 is responsible for overall control of the discriminator generation device 1, and includes an initial setting unit 101, an identification score calculation unit 102, a weak classifier learning unit 103, an identification weight calculation unit 104, and a discriminator generation unit 105. And an error weight calculation unit 106. The function of the control unit 10 is realized by, for example, a CPU (Central Processing Unit) developing and executing a program stored in the storage unit 13 of the discriminator generation device 1 in a memory unit (not shown).

The initial setting unit 101 initializes weak classifiers h _{t, k} , identification weights α _t , error weights D _{t, i, k} , and identification scores r _{t, k} that are variables. Specifically, appropriate values are set in the parameters of the weak classifiers _{ht, k} for each weak classifier shown in the related art described later. Real values are set for the identification weight α _t and the identification score r _{t, k} . For the identification score r _{t, k} , all error weights D _{t, i, k} are set to 1 / (NK). These initial values are input as appropriate values for the discriminator generation device 1 to execute each calculation process described later in the time frame t = 1.
Further, the initial setting unit 101 acquires learning data A via the input unit 11. Then, the initial setting unit 101 stores each constant related to the learning data A (the total number N of learning data, the total number K of classes, the total number T of time frames, the number of dimensions d of the time series signal x _i ) as constant information 131. Store in the unit 13.

The identification score calculation unit 102 calculates the identification score r _{t, k according} to the following (Equation 1) for all classes k (k = 1, 2,..., K). Then, the identification score calculation unit 102 stores the calculated identification score r _{t, k} as the identification score information 135 in the data storage unit 130.

Here, the function g _k (y _i ) is a variable that receives a correct class (class variable) y _i and returns “1” if y _i = k, and “−1” otherwise. This discrimination score r _{t, k} is a question of whether the correct class y _i and the discrimination result h _{t, k} (x _{i, t} ) by the weak discriminator h _{t, k} belong to the class k or not. If they match, the value increases. The identification score r _{t, k} increases as the error weight D _{t, i, k} described later increases.

The weak classifier learning unit 103 performs, for all classes k (k = 1, 2,..., K), existing weak classifiers such as decision stumps, kNN classifiers, linear Fisher classifiers (for example, “Bishop, C M., “Pattern Recognition and Machine Learning”, Springer Japan, 2007 ”), the weak classifiers h _{t, k} (t = 1, 2,..., T, k = 1, 2,. , K) are set so as to satisfy the following (Equation 2). The weak classifier learning unit 103 stores the set weak classifier h _{t, k} as weak classifier information 132 in the data storage unit 130.

This (Equation 2) means that the parameters of the weak classifier h _{t, k} are determined so as to maximize the discrimination score r _{t, k} .

Here, parameters constituting the weak classifier h _t, and _k is weak classifier h _{t used,} is defined by the type of _k. For example, in the case of decision stump, the dimension and threshold value for setting a threshold, in the case of a kNN discriminator, a specific level distribution, in the case of a linear Fisher discriminator, a direction vector (eigenvector) of the discriminant axis, It is the eigenvalue and the threshold on the discrimination axis. Note that any weak classifier h _{t, k} can be used as long as it is a known classifier. Further, in the actual calculation, the setting for maximizing the discrimination score r _{t, k} may not take the true maximum value, and the discrimination score r is more than the default (initial) weak discriminator h _{t, k.} Anything may be used as long as _{t and k} are set to approach the maximum value.

The identification weight calculation unit 104 uses the identification score r _{t, k} calculated by the identification score calculation unit 102 to indicate the reliability of the weak classifiers _{ht, k} of all classes set for each time frame. _t is calculated by the following (formula 3). Then, the identification weight calculation unit 104 stores the calculated identification weight α _t as identification weight information 133 in the data storage unit 130.

As the identification weight α _t calculated by (Equation 3), the identification score (Σ _k rt _{, k} ) for all K classes is used. As a result, the time-series signal discriminating apparatus 2 determines the weight of the weak discriminator _{ht, 1: k} for each time frame by the discriminating weight α _t considering the discriminating ability of all the classes at the time of discriminating the discrimination target data. It is possible to generate the strong discriminator H _k by setting optimally.

  The discriminator generation unit 105 determines whether or not the learning data acquired in the oldest time frame satisfies t <T (that is, whether or not the data of the last time frame T has been processed), and t <T If so, the data is transferred to the error weight calculation unit 106. When t = T, it is assumed that all learning data A has been processed, and constant information 131, weak classifier information 132, and identification weight information 133 stored in the data storage unit 130 are stored in the classifier 100.

The error weight calculation unit 106 updates the error weight D _{t, i, k} using (Equation 4) below. Then, the time frame t is incremented by 1 and delivered to the identification score calculation unit 102.

Here, ∝ means normalization so that Σ _{i, k} D _{t + 1, i, k} = 1. This (Expression 4) represents that the weight is decreased if the recognition result by the weak classifier _{ht, k} is correct, and is increased if the recognition result is incorrect. Therefore, when setting the weak classifier _{ht, k in} the next time frame, the recognition error is corrected so as to correctly recognize the sequence that failed to be recognized in this time frame. By repeating this, the weak classifiers _{ht, k} and the identification weight α _t are set so as to correct as many recognition errors as possible in each time frame, so that early recognition is realized.

(Operation procedure)
Next, an operation procedure of the discriminator generation device 1 according to the present embodiment will be described along FIG. 2 with reference to FIG.
FIG. 2 is a flowchart showing a process flow of the discriminator generation device 1 according to the present embodiment.

First, the initial setting unit 101 of the discriminator generation device 1 performs, as preprocessing for inputting the learning data A, a weak discriminator h _{t, k} , an identification weight α _t , an error weight D _{t, i, k} , and an identification score r _{t. , k} is set to an initial value (step S101). Specifically, an appropriate value is set as each weak classifier _{ht, k} for each parameter of the weak classifier _{ht, k} . Appropriate real values such as “0” and “1” are set for the identification weight α _t and the identification score r _{t, k} . For the identification score r _{t, k} , all error weights D _{t, i, k} are set to 1 / (NK). Then, the initial setting unit 101 stores these set numerical values in the data storage unit 130.

Next, the initial setting unit 101 acquires the learning data A via the input unit 11, and each constant related to the learning data A (the total number N of learning data, the total number K of classes, the total number T of time frames, the time series signal). The dimension number d) of x _i is stored in the constant information 131 of the data storage unit 130. Then, the initial setting unit 101 sets the time frame to t = 1 (step S102).

Subsequently, the identification score calculation unit 102 acquires the learning data A sequentially from the time frame t = 1 _, and calculates the identification score r _{t, k} using the above-described (Equation 1) (step S103). Then, the identification score calculation unit 102 stores the calculated identification score r _{t, k} in the identification score information 135 of the data storage unit 130.

Next, the weak discriminator learning unit 103 sets the weak discriminator h _{t, k} so as to maximize the discrimination score r _{t, k} calculated in step S103 according to (Equation 2) described above (step 2). S104). The weak classifier learning unit 103 stores the set weak classifier h _{t, k} in the weak classifier information 132 of the data storage unit 130.

Subsequently, the identification weight calculation unit 104 calculates the identification weight α _t using (Equation 3) described above (step S105). Then, the identification weight calculation unit 104 stores the calculated identification weight α _t in the identification weight information 133 of the data storage unit 130.

  Next, the discriminator generation unit 105 determines whether or not the time frame is t = T (step S106). If t <T instead of t = T (step S106 → No), it is determined that there is learning data A that has not yet been processed, and the process proceeds to the next step S107.

In step S107, the error weight calculator 106 updates the error weight D _{t, i, k} using the above-described (Equation 4). Then, the error weight calculation unit 106 stores the updated error weight D _{t, i, k} in the error weight information 134 of the data storage unit 130. Subsequently, the error weight calculation unit 106 increases the time frame t by one (step S108), returns to step S103, and continues the treatment.

  On the other hand, if it is determined in step S106 that t = T (step S106 → Yes), it is determined that all the learning data A has been processed, and the process proceeds to the next step S109. In step S 109, the classifier generation unit 105 acquires constant information 131, weak classifier information 132, and identification weight information 133 among the data stored in the data storage unit 130, and stores the information in the storage unit 13. The discriminator 100 is generated.

As described above, according to the discriminator generation device 1 and the discriminator generation method according to the present embodiment, t = T while increasing the value of the time frame t one by one for the time series signals composed of three or more classes. The process is repeated until it reaches _, and values such as the weak classifiers h _{t, k} and the identification weight α _t can be optimally set. Then, the set constant information 131, weak discriminator information 132, and discriminating weight information 133 are generated as the discriminator 100 and used by the time-series signal discriminating apparatus 2. The number of weak classifiers h _{t, k} and the identification score r _{t, k} and the like set by the identifier generating apparatus 1 does not increase only linearly with respect to time series length T and the number of classes K. Therefore, the time-series signal discriminating apparatus 2 uses this discriminator 100 to reduce the calculation load as compared to the method of configuring the two-class discriminator for all the class pairs using the binomial coefficient _k C ₂ described above. Can do.

  In this embodiment, the discriminator 100 generated by the discriminator generation device 1 is generated in the storage unit 13 of the discriminator generation device 1. However, the discriminator generation unit 105 is connected via the output unit 12. The data may be stored in a storage unit installed outside the discriminator generation device 1 or may be directly output to the time-series signal discrimination device 2 via a communication network or the like. Further, when the time-series signal identification device 2 is configured to include the classifier generation device 1, the time-series signal identification device 2 includes constant information processed by the classifier generation device 1 and stored in the storage unit 13. 131, weak classifier information 132, and identification weight information 133 can be acquired and processed.

≪Time-series signal identification device≫
Next, the time-series signal identification device 2 will be specifically described. FIG. 3 is a functional block diagram illustrating a configuration example of the time-series signal identification device 2 according to the present embodiment. The time-series signal identification device 2 is a device that identifies a class of input time-series signals (identification target data) using the classifier 100 generated by the classifier generation device 1 according to the present embodiment. As illustrated in FIG. 3, the control unit 20 includes an input unit (first input unit) 21, an output unit 22, and a storage unit 23.

First, the time series signal x input to the time series signal identification device 2 will be described. Unlike the above-described learning data A, the time series signal (identification target data) x = {x ₁ , x ₂ ,..., X _L } is assigned the class y∈ {1,. The time series signal identification device 2 identifies the class. Here, L is the sequence length of the identification target data, and L ≦ T. Incidentally, (in the case of L> T) If the sequence length L of the identification object data exceeds the sequence length T of the learning data A inputted to the discriminator generating apparatus 1, x T _{+ 1, ...,} corresponding to the data x _L Since the time-series signal discriminating apparatus 2 is not equipped with the weak discriminators h _{t, k and the} like, the processing is not performed and the processing is performed with L = T. The input time series signal may be a time series signal stored in advance in the storage unit 23 or the like, or may be a time series signal of a predetermined time frame acquired from a video camera or a microphone. Good.

Hereinafter, the configuration of the time-series signal identification device 2 will be specifically described with reference to FIG.
The input unit (first input unit) 21 includes an input interface for inputting information from the outside, receives an input of a time-series signal (identification target data) x, and passes it to the control unit 20. Further, the input unit 21 receives each piece of information of the discriminator 100 generated by the discriminator generation device 1 according to the present embodiment and stores it in the storage unit 23. Here, the information of the discriminator 100 acquired by the time-series signal discriminating apparatus 2 includes learning data index number N, class K, time frame T, dimension number d, and weak discriminator information stored as constant information 131. 132 and information regarding the identification weight information 133.

  The output unit 22 includes an output interface that outputs information generated by the time-series signal identification device 2, and outputs an identification result (identified class) of the time-series signal stored in the storage unit 23. The storage unit 23 stores the classifier 100 acquired from the classifier generation device 1 via the input unit 21 and the processing result of the identification target data as the identification result information 230.

  The control unit 20 controls the entire time-series signal identification device 2 and includes a strong classifier generation unit 201 and a data identification unit 202. The function of the control unit 20 is realized, for example, when the CPU develops and executes a program stored in the storage unit 23 of the time-series signal identification device 2 in a memory unit (not shown).

The strong classifier generation unit 201 acquires information on the classifier 100 via the input unit 21 and stores the information in the storage unit 23. The strong classifier generation unit 201, the weak classifier h _t that is stored in the identifier _100, by using the _k and the identification weight alpha _t, weak classifier h _t at each time _frame, the _k, (Equation 5 ) To generate strong classifiers H _k (class k = 1, 2,..., K) integrated for each class so as to satisfy the above.

Note that strong classifier generation unit 201, when the identification weight alpha _t the discriminator 100 is not stored, without using the identification weight alpha _t, weak classifier h _t identified as belonging to its own _{class, k} of the identification result, the weak classifier has been identified as not belonging to its own class h _t, and set as a value higher than the identification result of _k, the time frame each identification result of the weak classifier h _{t, k} for each class Can be added to generate a strong classifier H _k (class k = 1, 2,..., K).
Further, when the strong classifier H _k is generated as shown in (Equation 5) using the classification weight α _t stored in the classifier 100, the weight of the weak classifier _{ht, k} is set for each time frame. It is possible to generate a strong classifier H _k for each class by optimally setting. Therefore, it is possible to further improve the identification accuracy of the time-series signal identification device 2.

If it is considered that a sufficiently reliable identification result is obtained, the identification target data received by the input unit 21 is x ′ = {x_1, x_2,..., X_L ′} (L ′ ≦ L) and then may output the identification result before by using the identifying strong classifier H _k which is a L = L ', where L frames of the time-series signal is inputted in (equation 5).

The data discriminating unit 202 receives the time series signal x via the input unit 21 and uses the strong discriminator H _k for each class generated by the strong discriminator generating unit 201 to indicate how much the time series signal x is class k. Calculate how.

  Then, the data identification unit 202 calculates the identification result yε {1, 2,..., K} of the time series signal (identification target data) x so as to satisfy the following (Expression 6).

This (Equation 6), the strong class classifier H _k indicating the maximum value among the calculation results by the strong classifier H _k of each class of strong classifier generation unit 201 in (Equation 5) is generated, is input This means that the time series signal (identification target data) x belongs to the class.

  Further, the data identification unit 202 stores the calculation result (identified class y) in the storage unit 23 as identification result information 230, and the identification result information 230 is output to the external device (for example, a display) via the output unit 22. Output to.

(Operation procedure)
Next, an operation procedure of the time-series signal identification device 2 according to the present embodiment will be described along FIG. 4 with reference to FIG.
FIG. 4 is a flowchart showing a processing flow of the time-series signal identification device 2 according to the present embodiment.

  First, the strong discriminator generating unit 201 of the time series signal discriminating apparatus 2 acquires each piece of information (constant information 131, weak discriminator information 132, discriminating weight information 133) of the discriminator 100 through the input unit 21. It memorize | stores in the memory | storage part 23 (step S201).

The strong classifier generation unit 201, the weak classifier h _t that is stored in the identifier _100, by using the _k and the identification weight alpha _t, weak classifier h _t at each time _frame, the _k, (Equation 5 ) To generate strong classifiers H _k (class k = 1, 2,..., K) integrated for each class so as to satisfy () (step S202).

Next, the data identification unit 202 selects one of the classes kε {1, 2,..., K} that has not yet been calculated using the strong classifier H _k (x) (step S203). .

Subsequently, the data identification unit 202 acquires the identification target data x via the input unit 21, and calculates the strong classifier H _k (x) of the class selected in step S203 using (Equation 5). (Step S204).

Then, the data identification unit 202 determines whether or not the strong classifier H _k (x) has been calculated for all classes k = {1, 2,..., K} (step S205). If there are still classes that have not been calculated yet (step S205 → No), the process returns to step S203 to continue the processing. On the other hand, when the calculation for all classes k is completed (step S205 → Yes), the process proceeds to the next step S206.

  In step S206, the data identification unit 202 calculates the identification result yε {1, 2,..., K} of the time series signal (identification target data) x using the above-described (Equation 6) (step S207). . Then, the data identification unit 202 stores the class y of the time series signal (identification target data) x, which is the calculation result, in the storage unit 23 as identification result information 230 (step S207).

By doing in this way, according to the time series signal identification device 2 and the time series signal identification method according to the present invention, it becomes possible to efficiently recognize the time series signals of three or more classes efficiently. In addition, the weak classifier h _{t, k} included in the classifier 100 generated by the classifier generation apparatus 1 uses the error weight D _{t, i, k} to correct the recognition error in each time frame. Therefore, the time-series signal identification device 2 according to the present invention can further improve the early recognition performance.

<< Experimental Example of the Present Invention >>
Next, an experimental example using the time-series signal identification device 2 according to the present invention will be described. The experiment used time-series signals related to the driver's movements taken by a camera installed in the drive simulator. Seven drivers are allowed to perform 30 sets of driving simulations, in which each driver is instructed to perform 12 classes of operation. For example, in the class, operations such as “handle holding”, “air conditioner operation”, and “car navigation operation” are set. The positions of the left hand, right hand, and left elbow of each driver are automatically tracked from the moving image obtained by capturing the driver's motion, and the tracking result is used as identification target data (time-series signal). Since the identification target data (time series signal) is composed of the image coordinates (h, w) of the three joints, it becomes a time series signal of a d = 6 dimensional real vector.

In addition, a decision stump is used as the weak classifier h (see “pattern recognition and machine learning” described above). This decision stump weak classifier h is the simplest binary classifier. For the d-dimensional vector x _{i, t} , one of the d vector elements is selected, and a class label {1, −1} is given by threshold determination for the vector element. The parameters of the weak classifier h are set as follows, for example.
1. j ∈ {1, 2, ..., d}
2. v ∈ R
3. s ∈ {1, -1}
Here, j is a dimension index used for identification. v is an identification threshold in the specified dimension. Further, s determines whether the class is positive or negative, and is set, for example, as in (Expression 7) and (Expression 8) below.

  Experiments were performed by preparing learning data with the correct class for this identification target data (time-series signal). In the experiment, the cross-validation method (6 intersections) was performed to calculate the class recognition performance.

FIG. 5 shows an average identification rate of time series signals obtained by seven drivers using the time series signal identification apparatus 2 according to the present embodiment. The horizontal axis represents the time frame of the identification target data, and the vertical axis represents the identification rate. The thick line represents the recognition rate of the time-series signal discriminating apparatus 2 according to the present embodiment, and the thin line _represents the update processing of the error weights D _{t, i, k} important for early recognition (the error of the discriminator generating apparatus 1 according to the present embodiment). This is the recognition rate when the weak classifier h is set without performing the process of (Equation 4) by the weight calculation unit 106. As is apparent from the results shown in FIG. 5, it is shown that the performance relating to the recognition rate and early recognition is improved by updating the error weights D _{t, i, k} .

FIG. 6 is a diagram comparing the time-series signal identification device 2 according to the present embodiment and the Earlyboost (early recognition boosting) method which is an existing early recognition method. Here, in order to handle the Earlyboost method with more than three classes, we prepared classifiers that solve the two-class identification problem of “classes of interest” and “other classes” for each of the K classes. The method of comparing the identification scores was used. As shown in FIG. 6, it can be seen that the time-series signal discriminating apparatus 2 according to the present embodiment improves the discrimination rate more accurately and earlier than when the Earlyboost method is applied to multiple classes. . In the discriminator generating device 1 according to the present embodiment shown in FIG. 5, even when the time series signal discriminating device 2 performs the treatment without performing the update processing of the error weights D _{t, i, k.} As compared with the case where the Earlyboost method shown in FIG.

DESCRIPTION OF SYMBOLS 1 Classifier production | generation apparatus 2 Time-sequential signal identification apparatus 10,20 Control part 11,21 Input part 12,22 Output part 13,23 Storage part 100 Classifier 101 Initial setting part 102 Discrimination score calculation part 103 Weak classifier learning part 104 Discrimination weight calculation unit 105 Discriminator generation unit 106 Error weight calculation unit 130 Data storage unit 131 Constant information 132 Weak discriminator information 133 Discrimination weight information 134 Error weight information 135 Discrimination score information 201 Strong discriminator generation unit 202 Data discrimination unit 230 Discrimination Result information

Claims (11)

A time-series signal identification device that identifies a time-series signal, which is a sequence in which identification target data in the time frame are arranged, for each time frame obtained by dividing the time-series signal at predetermined time intervals, into three or more classes,
A first input unit that receives an input of a sequence of identification target data from the beginning of the time frame to a predetermined time frame shorter than the time-series signal;
A strong classifier generating unit that generates, for each class, a strong classifier that calculates the degree to which the series of identification target data received by the first input unit belongs to the class;
Using the strong classifier for each class generated by the strong classifier generation unit, the degree to which the classification target data received by the first input unit belongs to the class is calculated, and is calculated for each class. A classifier of the strong discriminator that indicates the maximum value of the degree belonging to the class, and a data discriminating unit to which the time series signal belongs,
A time-series signal identification device that identifies a class to which the time-series signal belongs without identifying all the time-series signals. The first input unit includes a weak classifier that identifies whether or not identification target data prepared for each time frame and each class belongs to its own class, and a weak classifier in each time frame And an identification weight representing the reliability of
The strong classifier generation unit, for each class, the weak classifier for each time frame from the beginning of the time frame received by the first input unit to a predetermined time frame shorter than the time series signal, The time-series signal discriminating apparatus according to claim 1, wherein the strong discriminator is obtained by adding the discrimination weights of the weak discriminators. Second input for accepting input of learning data in which each time frame is associated with a time-series signal as learning target data for the time frame and a correct class to which the learning target data belongs. And
The time series signal of each time frame in the learning data received by the second input unit is acquired in order of oldest of the time frame, the time series signal of the acquired time frame, the identification result of the weak classifier, An identification score calculation unit that calculates an identification score in the time frame for each of the classes, so that the value increases as the correct class to which the time-series signal of the time frame in the learning data belongs is matched,
A weak classifier learning unit that sets the weak classifier for each time frame and for each class so as to increase the calculated identification score;
An identification weight setting unit that sets an identification weight for each time frame so that the value increases as the identification score for each time frame and for each class increases.
The weak classifier set by the weak classifier learning unit and the identification weight set by the identification weight setting unit are used as the weak classifier and the identification weight input to the first input unit. The time-series signal identification device according to claim 2. A result of identifying the time series signals acquired in the oldest order of the time frames by the weak classifier set by the weak classifier learning unit, and a correct class to which the time series signals of the time frames indicated in the learning data belong The error weight is calculated so that the value becomes smaller if the weak classifier identification result is correct, and the value becomes larger if the weak classifier identification result is incorrect, and the calculated error is calculated. An error weight calculator configured to set a weight as the error weight for the learning data of the time frame next to the time frame in the acquired time-series signal;
The time series according to claim 3, wherein the identification score calculation unit calculates the value of the identification score in a time frame of learning data having a large error weight calculated by the error weight calculation unit. Signal identification device. The identification score calculation unit
As a result of identifying the time series signal of the acquired time frame by the weak classifier, when the time series data not belonging to the class can be correctly identified as not belonging to the class, the identification score in the class and the time frame The time-series signal identification device according to claim 4, wherein the identification score is calculated so that the value becomes higher. A time-series signal used in a time-series signal identification device that identifies a time-series signal, which is a series in which identification target data in the time frame are arranged, for each time frame obtained by dividing the time-series signal at predetermined time intervals into three or more classes. An identification method,
The time-series signal identification device is
Receiving an input of a series of identification target data from the beginning of the time frame to a predetermined time frame shorter than the time series signal;
Generating for each class a strong classifier that calculates a degree to which the received series of identification target data belongs to the class;
Using the generated strong classifier for each class, calculate the degree to which the received series of identification target data belongs to the class, and indicate the maximum value of the degrees belonging to the class calculated for each class Performing a step of setting a class of the strong classifier to a class to which the time series signal belongs,
A time-series signal identification method characterized by identifying a class to which the time-series signal belongs without identifying all of the time-series signals. A weak classifier for identifying whether or not identification target data prepared for each time frame and for each class belongs to its own class, and an identification weight representing the reliability of the weak classifier in each time frame Execute the step that accepts input,
In the step of generating the strong classifier for each class,
For each class, the weak classifier for each time frame from the beginning of the received time frame to a predetermined time frame shorter than the time-series signal is added by the identification weight of the weak classifier. The step of making a strong classifier
The time-series signal identification method according to claim 6, further executed. Receiving for each time frame an input of learning data in which a time-series signal as identification object data for learning of the time frame and a correct class to which the identification object data for learning belong are associated;
The time-series signal of each time frame in the received learning data is acquired in the oldest order of the time frame, and the time-series signal of the acquired time frame is indicated in the identification result of the weak classifier and the learning data. Calculating an identification score in the time frame for each of the classes such that the value increases as the correct class to which the time-series signal of the time frame belongs matches;
Setting the weak classifier for each time frame and for each class so as to increase the calculated identification score;
Further executing the step of setting an identification weight for each time frame so that the value increases as the identification score for each time frame and for each class increases.
The time according to claim 7, wherein the set weak discriminator and the set discriminating weight are the weak discriminator and the discriminating weight input in the step of receiving the input of the learning data. Sequence signal identification method. Compare the result of the weak classifier identifying the time-series signal acquired in the oldest order of the time frame with the correct class to which the time-series signal of the time frame indicated in the learning data belongs, and The error weight is calculated so that the value becomes small if the discrimination result of the classifier is correct, and the value becomes large if the discrimination result of the weak classifier is incorrect, and the calculated error weight is obtained. Further executing the step of setting as the error weight for the learning data of the next time frame of the time frame in the time series signal,
The time series signal identification method according to claim 8, wherein in the step of calculating the identification score, the value of the identification score in a time frame of learning data having a large calculated error weight is calculated to be large. . In the step of calculating the identification score,
As a result of identifying the time series signal of the acquired time frame by the weak classifier, when the time series data not belonging to the class can be correctly identified as not belonging to the class, the identification score in the class and the time frame The time-series signal identification method according to claim 9, wherein the identification score is calculated so as to increase.   The program for making a computer perform the time series signal identification method of any one of Claim 6 thru | or 10.

Images