CN116453337A - Machine learning-based vehicle driving behavior prediction method - Google Patents

Abstract

The present application discloses a method for predicting vehicle driving behavior based on machine learning. The method constructs a driving behavior chain and a vehicle state chain respectively by acquiring historical driving behavior data and historical vehicle motion state data of the target driver and the target driving vehicle. And the hidden Markov model is used to predict the state, and then predict the behavior of the target driver and the state of the target driving vehicle. By inputting the state prediction results into the preset route prediction model, the predicted route of the target driving vehicle is obtained and provided to the remote intelligent traffic control center for traffic control. The method can effectively improve driving safety and reduce traffic accident rate.

Description

A Machine Learning Based Vehicle Driving Behavior Prediction Method

technical field

The present application relates to the technical field of vehicle control, in particular to a method for predicting vehicle driving behavior based on machine learning.

Background technique

With the acceleration of urbanization and the popularization of transportation means, urban traffic congestion has become a global problem. Traffic congestion not only affects the normal operation of urban traffic, but also directly affects people's travel and quality of life. At present, the solutions to urban traffic congestion mainly include expanding the scale of roads, increasing public transport systems, and optimizing traffic management. However, there are some problems in these schemes, such as the expensive cost of expanding the road scale, long construction period, and easy to cause environmental pollution and other problems; increasing the public transportation system requires a large amount of capital investment, and needs to establish a complete road network and station facilities, etc., which is difficult It will be effective in the short term; optimizing traffic management requires a large amount of artificial intelligence technology and data analysis methods, and requires a large amount of data support, and the reliability and effect need to be improved.

In this case, intelligent traffic control technology based on machine learning has gradually become an area of great concern. By utilizing a large amount of historical data and real-time data, intelligent traffic control technology based on machine learning can predict traffic flow, predict traffic congestion, adjust traffic lights, optimize routes, etc., thereby improving the efficiency and quality of urban traffic.

At present, there are many relevant patent documents disclosing intelligent traffic control technology based on machine learning. For example, U.S. Patent US20190377550A1 discloses a traffic congestion control method based on deep reinforcement learning, which uses deep reinforcement learning algorithms to predict traffic congestion, thereby achieving traffic congestion control. This method adopts an unsupervised learning method, that is, automatically learns the traffic congestion of road sections, so as to adjust the timing of signal lights to optimize traffic flow.

However, there are still some problems in the current intelligent traffic control technology. First of all, existing technologies often only rely on real-time data such as vehicles and road networks, but do not make full use of information such as driver behavior and status. This will cause the control system to be unable to fully and accurately understand and predict traffic conditions, which limits the effect of intelligent traffic control. Secondly, the existing technology can usually only achieve a single function, such as adjusting the timing of signal lights and optimizing routes.

To address these limitations, the researchers proposed a machine learning-based driving behavior prediction method. For example, US patent application 20190355862A1 introduces a deep learning-based driving behavior prediction system and method, which uses neural networks to learn patterns of driving behavior and predicts driving behavior based on dynamic factors such as driving environment, vehicle state, and road conditions. Similarly, US patent application 20190110392A1 introduces a system and method for vehicle driving behavior recognition based on deep learning, which uses neural networks to learn the characteristics of driving behavior, and classifies and recognizes driving behavior based on the behavior characteristics.

However, some problems still exist in this prior art. For example, some methods only focus on a single driving behavior, while ignoring the interrelationships and influences between driving behaviors, so they cannot accurately predict the driving behavior during the entire driving process; some methods only consider a small number of factors in training and prediction, Such as vehicle speed and acceleration, etc., without considering other factors that may affect driving behavior, such as the driver's psychological state and behavior habits. In addition, some methods in the prior art may require a large amount of data to train the model, and the time for training and prediction is long, which is not suitable for real-time application scenarios.

Therefore, there is a need for a new driving behavior prediction method based on the interaction between multiple factors and driving behaviors, which can accurately predict the driver's driving behavior and provide it to the intelligent traffic control system to achieve safer, Efficient and smart traffic management.

Contents of the invention

The embodiment of the present application provides a method for predicting vehicle driving behavior based on machine learning. The method can obtain the real-time driving behavior data of the target driver in real time, and use the Markov model to predict the behavior and state of the target driver, so as to predict the driving behavior of the target driver. future behavior and state. At the same time, the patent of the present invention provides an intelligent traffic control method based on a route prediction model, which can use the predicted route to plan and optimize traffic flow in advance, thereby improving traffic efficiency and safety.

The embodiment of the present application provides a method for predicting vehicle driving behavior based on machine learning, including:

A method for predicting vehicle driving behavior based on machine learning, said method comprising: determining a prediction target, said prediction target comprising: a target driver and a target driving vehicle; obtaining historical driving behavior data of the target driver and the history of the target driving vehicle Vehicle motion state data; based on historical driving behavior data, construct a driving behavior chain; the driving behavior chain is a tree structure, composed of nodes and connecting lines, each node represents a driving behavior, in the driving behavior chain, from the root node From the beginning to the end of each terminal node, a complete single driving behavior chain is formed, which characterizes all the driving behaviors performed by the target driver at the end of a complete vehicle driving; based on historical vehicle state data, a vehicle state chain is constructed; The vehicle state chain is a tree structure consisting of nodes and connecting lines. Each node represents a vehicle state. In the vehicle state chain, a complete single vehicle state is formed from the root node to the end node. chain, which characterizes all the vehicle states experienced by the target driving vehicle at the end of a complete vehicle driving; obtain the real-time driving behavior data of the target driver, and predict the driving behavior of the driver on the basis of the driving behavior chain, Obtain the behavior prediction result, obtain the real-time vehicle motion state data of the target driving vehicle, predict the vehicle state of the target driving vehicle on the basis of the vehicle state connection, and obtain the state prediction result; use the state prediction result as an input variable and input it to the prediction The established route prediction model obtains the predicted route of the target driving vehicle and provides it to the remote end for intelligent traffic control; the behavior prediction result is used as an input variable and input into the preset behavior judgment model to obtain the driving safety index of the target driver. It is provided to the remote end for intelligent traffic warning.

Further, the driving behavior data at least includes: seat belt wearing status, mobile phone use times, eating and drinking times and blinking times; the historical driving behavior data is the historical driving behavior data of the target driver; the real-time driving behavior data is Real-time driving behavior data of target drivers.

Further, the vehicle motion state data at least includes: position, speed, acceleration and steering angle; the historical vehicle motion state data is historical vehicle motion state data of the target driving vehicle; the real-time vehicle motion state data is real-time The vehicle motion state data of the target driving vehicle.

Further, the method for constructing a driving behavior chain based on historical driving behavior data includes: viewing the driving behavior chain as a sequence set composed of multiple driving behavior sequences; The driving behavior data is regarded as a driving behavior sequence; each historical driving behavior data contained in the driving behavior sequence is regarded as a state of driving behavior, and the corresponding driving behavior at the next moment is regarded as an observation of driving behavior value; consider the driving behavior sequence as a state sequence, where each driving behavior state is represented as a hidden state, and each driving behavior observation value is represented as an observed state; use a hidden Markov model to infer all possible The sequence of implicit states acts as a chain of driving behavior.

Further, the method of obtaining the real-time driving behavior data of the target driver, predicting the driving behavior of the driver on the basis of the driving behavior chain, and obtaining the behavior prediction result includes: in the driving behavior chain, according to the current driving behavior State and transition matrix, calculate the probability distribution of the next driving state; express the current state as a probability vector, and then calculate the probability distribution of the next state through vector multiplication; according to the calculated probability distribution of the next state, select The state with the highest probability is used as the next driving state of the target driver, find the state with the highest probability in the probability distribution, and then use it as the next state; repeat the above steps until the end of the predicted time period; during this process, record The driving state of the target driver at each time point, so as to obtain the behavior sequence as the prediction result of the target driver's future behavior.

Further, the method for constructing a vehicle state chain based on historical vehicle motion state data includes: converting historical vehicle motion state data into a discretized state; The value is transformed into the corresponding discrete state; the vehicle state sequence is regarded as a hidden Markov model, assuming that the vehicle state sequence s ₁ , s ₂ ,..., s _T is a first-order Markov chain with a length of T, where s _t represents the vehicle state at time t; estimate model parameters A, B and π, where A is the state transition matrix, B is the observation probability matrix, and π is the initial state probability vector; use historical vehicle motion state data to estimate model parameters, so that the model The likelihood value on the observed state sequence is the largest; use the model parameters and real-time vehicle motion state data to predict the vehicle state; use the forward algorithm to calculate the observed state at the current moment o ₁ , o ₂ ,..., o _t Under the condition, the driver’s state is the posterior probability P(s _t =i|o ₁ , o ₂ ,..., o _t , A, B, π), and then the vehicle at the current moment is determined according to the maximum posterior probability criterion status, namely:

Using the predicted vehicle state sequence, the continuous same state is merged into one state node, thus constructing the vehicle state chain.

Further, the method of obtaining the real-time vehicle motion state data of the target driving vehicle, predicting the vehicle state of the target driving vehicle on the basis of the vehicle state chain, and obtaining the state prediction result includes: in the vehicle state chain, according to the current The vehicle state and transition matrix, calculate the probability distribution of the next vehicle state; express the current state as a probability vector, and then calculate the probability distribution of the next state through vector multiplication; according to the calculated probability distribution of the next state , select the state with the highest probability as the next vehicle state of the target driving vehicle, find the state with the highest probability in the probability distribution, and then use it as the next state; repeat the above steps until the end of the predicted time period; in this process, Record the vehicle state of the target driving vehicle at each time point, so as to obtain the vehicle state sequence as the prediction result of the future behavior of the target driving vehicle.

Further, the method of using the state prediction result as an input variable and inputting it into the preset route prediction model to obtain the predicted route of the target driving vehicle, and providing it to the remote end for intelligent traffic control includes: obtaining historical route data of the target driving vehicle and the corresponding historical state results are trained using machine learning to obtain a route prediction model, the input variable of the route prediction model is the state prediction result, and the output result is the predicted route.

Further, the method of inputting the behavior prediction result as an input variable into the preset behavior judgment model to obtain the driving safety index of the target driver and providing it to the remote end for intelligent traffic warning includes: assuming that the behavior prediction result corresponds to The driving behavior sequence is S=s ₁ , s ₂ ,...,s _n , where s _i represents the driving behavior state of the target driver at time point i, and this sequence is input into the behavior judgment model to obtain the driving safety index S _c , The formula is as follows:

Among them, N is the length of the behavior sequence, M _i is the number of all possible driving behaviors at time point i, p(j|i) is the probability of predicting the jth driving behavior at time point i, weight _j is the The weight of j kinds of driving behaviors; the judgment model first predicts possible driving behaviors and corresponding probabilities for each time point, and then for each time point, calculates the weighted average value of all possible driving behaviors, wherein, the higher the weight Larger means more dangerous, and finally the average value of all time points is used as the driving safety index.

Further, the remote end is a remote intelligent traffic control center.

A method for predicting vehicle driving behavior based on machine learning provided by the application has the following beneficial effects:

First of all, the patent of the present invention provides a method for constructing a driving behavior chain, which can use the historical behavior data of the target driver or vehicle to construct a behavior chain model with a certain time series relationship. This model can be predicted by the Markov model and used in route prediction and behavior judgment, with high accuracy and reliability. Compared with the traditional traffic management method, the intelligent traffic method based on the driving behavior chain of the patent of the present invention can better reflect the driving behavior of the target driver or vehicle, thereby more accurately predicting future behavior and improving driving safety.

Secondly, the patent of the present invention provides a route prediction method based on the prediction results, which can use the Markov model to predict the future driving state of the target driving vehicle, and use the prediction results for route prediction. This approach enables more accurate prediction of the travel route of the target driving vehicle, improving the efficiency of the traffic method. At the same time, the route prediction results can also be used as input variables for further behavior judgment and prediction, thereby further improving driving safety.

Finally, the patent of the present invention provides a behavior judgment method based on driving behavior chain and route prediction, which can use the driving behavior chain and route prediction results as input variables, and use machine learning algorithms to judge the behavior of the target driver. This method can more accurately predict the driving behavior of the target driver and provide the corresponding driving safety index, which can be used for remote intelligent traffic warning. The implementation of the method can more effectively reduce the incidence of traffic accidents and improve the safety and efficiency of traffic methods.

Description of drawings

The technical solutions and other beneficial effects of the present application will be apparent through the detailed description of the specific embodiments of the present application below in conjunction with the accompanying drawings.

FIG. 1 is a schematic structural diagram of a machine learning-based vehicle driving behavior prediction method provided by an embodiment of the present invention;

2 is a schematic structural diagram of a driving behavior chain or a vehicle state chain of a machine learning-based vehicle driving behavior prediction method provided by an embodiment of the present invention;

3 is a structural schematic diagram of a single driving behavior chain in the driving behavior chain or a single vehicle state chain in the vehicle state chain in a machine learning-based vehicle driving behavior prediction method provided by an embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the application with reference to the drawings in the embodiments of the application. Apparently, the described embodiments are only some of the embodiments of the application, not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts belong to the scope of protection of this application.

Example 1

A method for predicting vehicle driving behavior based on machine learning, said method comprising: determining a prediction target, said prediction target comprising: a target driver and a target driving vehicle; obtaining historical driving behavior data of the target driver and the history of the target driving vehicle Vehicle motion state data; based on historical driving behavior data, construct a driving behavior chain; the driving behavior chain is a tree structure, composed of nodes and connecting lines, each node represents a driving behavior, in the driving behavior chain, from the root node From the beginning to the end of each terminal node, a complete single driving behavior chain is formed, which characterizes all the driving behaviors performed by the target driver at the end of a complete vehicle driving; based on historical vehicle state data, a vehicle state chain is constructed; The vehicle state chain is a tree structure consisting of nodes and connecting lines. Each node represents a vehicle state. In the vehicle state chain, a complete single vehicle state is formed from the root node to the end node. chain, which characterizes all the vehicle states experienced by the target driving vehicle at the end of a complete vehicle driving; obtain the real-time driving behavior data of the target driver, and predict the driving behavior of the driver on the basis of the driving behavior chain, Obtain the behavior prediction result, obtain the real-time vehicle motion state data of the target driving vehicle, predict the vehicle state of the target driving vehicle on the basis of the vehicle state connection, and obtain the state prediction result; use the state prediction result as an input variable and input it to the prediction The established route prediction model obtains the predicted route of the target driving vehicle and provides it to the remote end for intelligent traffic control; the behavior prediction result is used as an input variable and input into the preset behavior judgment model to obtain the driving safety index of the target driver. It is provided to the remote end for intelligent traffic warning.

Specifically, the construction of behavior chains and state chains is a key step. Through the analysis and processing of historical driving behavior data and historical vehicle status data, a tree structure containing multiple nodes and connecting lines can be constructed. Each node represents a driving behavior or vehicle state, while connecting lines represent the transition relations between these nodes. In this way, the complete behavior or state information of the target driver or driving vehicle can be obtained by traversing the tree structure, so as to make better predictions.

Machine learning technology is used in the present invention, and the prediction result is input into a preset model for processing to obtain a more accurate prediction result. Through continuous feedback and learning, the prediction model can be continuously optimized to improve the accuracy and reliability of the prediction.

In Figure 2 and Figure 3, numbers are nodes, and dotted lines are connecting lines of nodes. The solid line is a complete single vehicle state chain or a complete single driving behavior chain.

Example 2

On the basis of the previous embodiment, the driving behavior data at least includes: seat belt wearing status, mobile phone use times, eating and drinking times and blinking times; the historical driving behavior data is the historical driving behavior data of the target driver; The real-time driving behavior data is the real-time driving behavior data of the target driver.

The driving behavior data includes at least: seat belt wearing status, mobile phone use times, eating and drinking times and blinking times. These driving behavior data are obtained from the behavior of the target driver. The seat belt wearing status can be monitored in real time through sensors and other devices, and information such as the number of times mobile phones are used, the number of meals and the number of blinks can be obtained through image recognition, voice recognition and other technologies. for extraction and analysis.

The historical driving behavior data refers to the historical driving behavior data of the target driver. These data can be obtained from the target driver's previous driving records, such as from vehicle black boxes or other data recording devices. Historical driving behavior data is one of the key factors in building a driving behavior chain. By analyzing historical data, a driving behavior chain with rich driving behavior information can be constructed to provide reference and support for the prediction of real-time driving behavior data.

The real-time driving behavior data refers to the real-time driving behavior data of the target driver. These data can be obtained in real time through technologies such as sensors, cameras, and voice recognition, and then analyzed and predicted based on the driving behavior chain. By monitoring and predicting the driving behavior of target drivers in real time, it can provide real-time support and feedback for traffic control and early warning, thereby improving traffic safety and efficiency.

In a word, the present invention specifies the type and source of the driving behavior data, and combines historical driving behavior data and real-time driving behavior data to construct a driving behavior chain, which provides a basis for driving behavior prediction. This technology can provide more comprehensive and accurate information support for intelligent traffic control and early warning.

Example 3

On the basis of the previous embodiment, the vehicle motion state data at least includes: position, speed, acceleration and steering angle; the historical vehicle motion state data is the historical vehicle motion state data of the target driving vehicle; the real-time vehicle motion state data The motion state data is real-time vehicle motion state data of the target driving vehicle.

The vehicle motion state data at least include: position, speed, acceleration and steering angle. These vehicle motion state data can be monitored and collected in real time through sensors on the vehicle or other devices, such as the Global Positioning System (GPS). Vehicle motion state data such as position, speed, acceleration, and steering angle can reflect the real-time motion state of the target driving vehicle and provide basic data for the construction of the vehicle state chain.

The historical vehicle motion state data refers to the historical vehicle motion state data of the target driving vehicle. These data can be obtained from equipment such as vehicle black boxes, on-board recorders, or other data recording devices. Historical vehicle motion state data is one of the key factors in building a vehicle state chain. By analyzing historical data, a vehicle state chain with rich vehicle motion state information can be constructed to provide reference and support for the prediction of real-time vehicle motion state data.

The real-time vehicle motion state data refers to real-time vehicle motion state data of the target driving vehicle. These data can be obtained in real time through on-board sensors, cameras and other equipment, and then analyzed and predicted according to the vehicle status chain. By monitoring and predicting the vehicle motion state of the target driving vehicle in real time, it can provide real-time support and feedback for traffic control and early warning, thereby improving traffic safety and efficiency.

Example 4

On the basis of the previous embodiment, the method for constructing a driving behavior chain based on historical driving behavior data includes: viewing the driving behavior chain as a sequence set composed of multiple driving behavior sequences; All historical driving behavior data corresponding to the same moment are regarded as a driving behavior sequence; each historical driving behavior data contained in the driving behavior sequence is regarded as a state of driving behavior, and the driving behavior corresponding to the next moment is regarded as a state of driving behavior. is an observation value of driving behavior; the driving behavior sequence is regarded as a state sequence, where the state of each driving behavior is represented as a hidden state, and the observed value of each driving behavior is represented as an observed state; using Hidden Markov The model infers all possible sequences of hidden states as chains of driving behavior.

The method of constructing a driving behavior chain based on historical driving behavior data uses a hidden Markov model (Hidden Markov Model, HMM). This method regards the driving behavior chain as a sequence set composed of multiple driving behavior sequences. By treating all the historical driving behavior data corresponding to each same moment in the historical driving behavior data as a driving behavior sequence, the driving behavior sequence is regarded as A state sequence, and infer all possible hidden state sequences through the HMM model, and finally act as a driving behavior chain.

Specifically, this method regards each item of historical driving behavior data included in the driving behavior sequence as a state of driving behavior, and regards the corresponding driving behavior at the next moment as an observed value of driving behavior. Then, the driving behavior sequence is regarded as a state sequence, where the state of each driving behavior is represented as a hidden state, and the observed value of each driving behavior is represented as an observed state. The hidden Markov model can estimate the state transition probability and observed state probability of each driving behavior by learning the driving behavior sequence in the historical driving behavior data, and calculate all possible hidden state sequences through these probabilities, thus constructing the Driving behavior chain.

Assume that the driving behavior data sequence is S=s ₁ , s ₂ ,...,s _T , where each driving behavior s _t can take K possible states, that is, s _t ∈1, 2,...,K; let O=o ₁ , o ₂ ,..., o _T is an observation sequence, where each observation state o _t can take V possible values, that is, 0 _t ∈ 1, 2,..., V. The transition probability matrix is A=[a _ij ]K×K, where aij represents the probability of transitioning from state i to state j; the observation probability matrix is B=[b _jt ]K×V, where bjt represents the observed state j The probability of observing state t; the initial state probability vector is π=[π _i ], where π _i represents the probability that the initial state is i; the driving behavior at any moment is only dependent on the driving behavior at the previous moment, namely: P( s _t |s _t-1 , s _t-2 ,...,s ₁ )=P(s _t |s _t-1 ); in this way, a K×K matrix A is used to represent the transition from one state to another The probability. Specifically, let a _ij represent the probability of transitioning from state i to state j, then: Indicates that the sum of all transition probabilities starting from state i is 1. It is set that the observation state at any moment only depends on the driving behavior state at that moment, and is independent of the state and observation state at other moments, namely: P(o _t |s _t , o _t-1 , s _t-1 , o _t-2 , s _t-2 ,..., o ₁ , s ₁ )=P(o _t |s _t ); use a K×V matrix B to represent the probability of observing state t under state j . Specifically, let b _jt represent the probability of observing state t in state j, then B can be expressed as: B=[b _jt ] _K×V ; where b _jt can be estimated according to training data. Usually, we can use the maximum likelihood estimation method or Bayesian estimation method to calculate the value of b _jt . For example, in maximum likelihood estimation, the estimate of b _jt can be calculated as: /> Among them, [o _t == v] indicates whether the observed state o _t is equal to v.

In the above formula, we can see that A and B are two important matrices, which respectively represent the probability of transitioning from one state to another and the probability of observing a certain observed state in a certain state. The calculation of these two matrices requires the use of training data for parameter estimation, and the most commonly used methods are the Baum-Welch algorithm or the Forward-Backward algorithm.

Specifically, in the Baum-Welch algorithm, we first randomly initialize A, B and π, and then repeatedly perform the following two steps until convergence:

Expectation step: In this step, we use the current model parameters A, B and π, and use the training data to calculate the posterior probability P(s _t =i|O, A , B, π). This probability can be calculated using a forward algorithm and a backward algorithm.

Maximization step: In this step, we use the current posterior probability to calculate new model parameters A, B, and π that maximize the likelihood of the training data under these parameters. Specifically, we calculate new parameter values as follows:

π _i =P(s ₁ =i|O, A, B, π)

Where [o _t == v] indicates whether the observed state o _t is equal to v, Indicates to sum over all training samples.

Finally, the parameters A, B and π obtained by the Baum-Welch algorithm can be used to predict driving behavior and construct a behavior chain.

Example 5

On the basis of the previous embodiment, the method for obtaining the real-time driving behavior data of the target driver, predicting the driving behavior of the driver on the basis of the driving behavior chain, and obtaining the behavior prediction result includes: In , calculate the probability distribution of the next driving state according to the current driving state and transition matrix; express the current state as a probability vector, and then calculate the probability distribution of the next state through vector multiplication; according to the calculated next The probability distribution of the state, select the state with the highest probability as the next driving state of the target driver, find the state with the highest probability in the probability distribution, and then use it as the next state; repeat the above steps until the end of the predicted time period; In this process, the driving state of the target driver at each time point is recorded, and the behavior sequence is obtained as the prediction result of the target driver's future behavior.

After we obtain the parameters A, B and π of the model, we can use these parameters to predict the future driving behavior state, and build a driving behavior chain based on these states.

Specifically, for a given set of observation sequences O=(o ₁ , o ₂ ,..., o _T ), we can use the forward-backward algorithm (Forward-Backward Algorithm) to calculate the model parameters A, B and π Next, the posterior probability P(S|O, A, B, π) of the driving behavior state sequence S=(s ₁ , s ₂ , . . . , s _T ). This posterior probability represents the probability of the occurrence of the driving behavior state sequence S in the case of the observation sequence O. Specifically, the posterior probability can be calculated by the following formula:

Among them, P(O|S, A, B) represents the probability of the observation sequence O appearing in the case of a given driving behavior state sequence S, which can be calculated using the forward algorithm; P(S|π, A) represents When the initial state distribution π and the state transition probability matrix A are fixed, the probability of the driving behavior state sequence S can be calculated using a dynamic programming algorithm; The probability of the observation sequence O in the case of B and π can be calculated using the forward algorithm.

After obtaining the posterior probability of the driving behavior state sequence, we can select the most likely driving behavior state sequence according to these probabilities as the representation of the driving behavior chain. Specifically, we can use the Viterbi algorithm to calculate the most likely driving behavior state sequence under the model parameters A, B and π The core idea of the Viterbi algorithm is to record the maximum probability to each state and the maximum probability path to the state while calculating the driving behavior state sequence S, so as to be used when calculating the next state. Finally, the driving behavior chain can be expressed as: C=(c ₁ , c ₂ , . . . , c _K ), where c _k represents the number of occurrences of driving behavior state k.

After using the Baum-Welch algorithm to get the parameters A, B and π of the model, we can use these parameters to predict driving behavior.

Specifically, given an observation sequence O=(o ₁ , o ₂ ,..., o _T ), we can use the forward algorithm to calculate the probability P(O |A, B, π). This probability can be used to evaluate the fit of the model, and it can also be used to predict the future driving behavior state.

When predicting driving behavior, we need to predict the driving behavior state at T' moments in the future. Specifically, assuming that the current moment is t, we want to predict the driving behavior state from t+1 to t+T′. The forecasting process can be divided into two steps:

Driving behavior state prediction _: In _the prediction process, we need to predict the driving behavior state sequence S=(s _{t+ 1} , s _t+2 ,..., s _t+T′ ). This prediction process can be done using the forward algorithm. Specifically, we can use the forward algorithm to calculate the driving behavior from t+1 to t+T′ when the driving behavior state is s _t at the current moment t The posterior probability distribution P( _st+1 , _st+2 ,..., _st+T ′|o ₁ , o ₂ ,..., o _t , A, B, π) of the state. This posterior probability can be calculated by the following recursive formula:

Among them, α _{t, i} represent the probability of the occurrence of the observation sequence O=(o ₁ , o ₂ ,..., o _t ) when the driving behavior state is i at time t, which can be calculated using dynamic programming; Indicates the posterior probability of the driving behavior state k from t+1 to t+T′ when the driving behavior state is i at time t, which can also be calculated using dynamic programming.

via caculation We can obtain the posterior probability distribution of each driving behavior state from t+1 to t+T′, so that the driving behavior state at each moment can be determined according to the maximum posterior probability criterion, namely:

where l=1, 2, . . . , T'.

Driving behavior chain construction: After predicting the driving behavior state at T′ time in the future, we can connect these states to build a driving behavior chain from the current moment to the future moment. Specifically, we can merge consecutive identical driving behavior states into one driving behavior node using the following rules:

If two adjacent driving behavior states are the same, they can be merged into one driving behavior node;

If two adjacent driving behavior states are different, they cannot be merged into one driving behavior node, and need to be used as different driving behavior nodes respectively. Using the above rules, we can construct a driving behavior chain from the current moment to the future moment, which describes the evolution of driving behavior in the future moment.

Example 6

On the basis of the previous embodiment, the method for constructing a vehicle state chain based on historical vehicle motion state data includes: converting historical vehicle motion state data into discretized states; specifically, vehicle position, speed, acceleration and Continuous real values such as steering angles are transformed into corresponding discrete states; the vehicle state sequence is regarded as a hidden Markov model, assuming that the vehicle state sequence s ₁ , s ₂ ,..., s _T is a first-order Markov with length T Cove chain, where s _t represents the vehicle state at time t; estimate model parameters A, B and π, where A is the state transition matrix, B is the observation probability matrix, and π is the initial state probability vector; using historical vehicle motion state data to Estimate the model parameters to maximize the likelihood of the model on the observed state sequence; use the model parameters and real-time vehicle motion state data to predict the vehicle state; use the forward algorithm to calculate the observed state at the current moment as o ₁ , o ₂ ,..., o _t , the posterior probability P(s _t =i|o ₁ , o ₂ ,..., o _t , A, B, π) of the driver's state i, then according to the maximum posterior probability criterion To determine the current state of the vehicle, namely:

Using the predicted vehicle state sequence, the continuous same state is merged into one state node, thus constructing the vehicle state chain.

Based on the construction process of the driving behavior chain explained earlier, we can build a vehicle state chain by analogy. Specific steps are as follows:

Preprocessing: Convert historical vehicle motion state data into a discretized state. Specifically, we can convert continuous real values such as vehicle position, velocity, acceleration, and steering angle into corresponding discrete states, such as dividing the speed into three states: slow, medium, and fast, and dividing the steering angle into left, right, and Turn and go straight three states, etc.

Establishment of Markov model: analogy to the establishment process of the driving behavior chain, we can regard the vehicle state sequence as a hidden Markov model, assuming that the vehicle state sequence s ₁ , s ₂ ,..., s _T is a length T A first-order Markov chain, where s _t represents the vehicle state at time t.

Parameter estimation: use the Baum-Welch algorithm to estimate model parameters A, B and π, where A is the state transition matrix, B is the observation probability matrix, and π is the initial state probability vector. Specifically, we can use historical vehicle motion state data to estimate model parameters, so that the likelihood value of the model on the observed state sequence is maximized.

Vehicle state prediction: Using model parameters and real-time vehicle motion state data, the vehicle state can be predicted. Specifically, we can use the forward algorithm to calculate the posterior probability of the driver's state i under the condition that the observed state is o ₁ , o ₂ ,..., o _t at the current moment

P(s _t =i|o ₁ , o ₂ ,..., o _t , A, B, π), and then determine the vehicle state at the current moment according to the maximum a posteriori probability criterion, namely:

Vehicle state chain construction: Using the predicted vehicle state sequence, consecutive identical states can be merged into one state node, thereby constructing a complete vehicle state chain, which describes the evolution process of the vehicle state.

Example 7

On the basis of the previous embodiment, the method for obtaining the real-time vehicle motion state data of the target driving vehicle, predicting the vehicle state of the target driving vehicle on the basis of the vehicle state connection, and obtaining the state prediction result includes: In the state chain, the probability distribution of the next vehicle state is calculated according to the current vehicle state and transition matrix; the current state is expressed as a probability vector, and then the probability distribution of the next state is calculated by vector multiplication; according to the calculated The probability distribution of the next state, select the state with the highest probability as the next vehicle state of the target driving vehicle, find the state with the highest probability in the probability distribution, and then use it as the next state; repeat the above steps until the end of the predicted time period ; In this process, record the vehicle state of the target driving vehicle at each time point, so as to obtain the vehicle state sequence as the prediction result of the future behavior of the target driving vehicle.

Specifically, we can use the following formula for parameter estimation: initial state probability vector π _i : π _i =P(s ₁ =i);

State transition probability matrix A _ij :

Observation probability matrix B _jt :

Among them, P(st _t =i, st _t+1 =j|o ₁ , o ₂ ,..., o _T , A, B, π) is the time t when the model parameters and observation state sequence are known. State is i, the conditional probability of state j at time t+1;

P(s _t =i|o ₁ , o ₂ ,..., o _T , A, B, π) is the posterior probability that the state at time t is i when the model parameters and observed state sequence are known; P (s _t = j, o _t = t|o ₁ , o ₂ ,..., o _T , A, B, π) is when the model parameters and observation state sequence are known, the state at time t is j, and the observation The joint probability of the state being t; P(s _t = j|o ₁ , o ₂ ,..., o _T , A, B, π) is the state at time t when the model parameters and the observed state sequence are known. The posterior probability of j.

Vehicle state prediction: Analogous to the prediction process of driving behavior, we can use prediction algorithms to predict future vehicle states. Specifically, given the observed state o _t at the current moment t, we can use the existing model parameters A, B, and π, as well as the historical information of the previous observed state, to predict the vehicle state at k moments in the future. This process can be carried out using the forward algorithm, and the specific recursive formula is:

Among them, α _{k, i} represent the forward probability of the vehicle state at time k when the vehicle state is i, and the observed state is o ₁ , o ₂ ,..., o _k , P(o _k |s _k =i, A, B) is when the state is i, under the condition that the observed state is o _k , the probability that the observed state is o _k , P(s _k =i | s _k-1 =j, A) is the transition from state j to state probability of i.

Construction of vehicle state links: analogous to the construction process of driving behavior chains, we can connect the predicted vehicle state sequences to form a vehicle state chain. Specifically, we can regard the predicted vehicle state _sk at each moment as a node of the state chain, regard the transition relationship between states as the connection line of the state chain, and finally construct a tree-structured vehicle state connection .

In general, based on historical vehicle motion state data, the process of building a vehicle state chain is similar to the process of building a driving behavior chain, both of which use hidden Markov models for prediction and chain construction. The difference is that the driving behavior chain is constructed based on the driver's historical driving behavior data, while the vehicle state chain is constructed based on the vehicle's historical motion state data.

In the specific implementation, we need to preprocess the vehicle motion state data, such as removing noise, filtering, etc., to ensure the accuracy of the data. Then, we can use the aforementioned algorithms and formulas to obtain model parameters by training the model, and then perform vehicle state prediction and connection construction.

When predicting the state of the vehicle, we can predict the state of the vehicle in different time periods by setting different values of $k$ at the time of prediction. At the same time, in practical applications, we can also combine the real-time acquisition of vehicle motion state data to continuously update the model parameters to improve the accuracy of prediction.

It should be noted that the state nodes and connection lines in the vehicle state connection do not directly correspond to the vehicle's motion state and the transition relationship between states, but are constructed based on the model's prediction results of these information.

Example 8

On the basis of the previous embodiment, the method of using the state prediction result as an input variable and inputting it into the preset route prediction model to obtain the predicted route of the target driving vehicle and provide it to the remote end for intelligent traffic control includes: obtaining the target The historical route data of the driving vehicle and the corresponding historical state results are trained using machine learning to obtain a route prediction model, the input variable of the route prediction model is the state prediction result, and the output result is the predicted route.

The state prediction result is used as an input variable, input into the preset route prediction model, and the predicted route of the target driving vehicle is obtained, which is provided to the remote end for intelligent traffic control. The machine learning method is used for route prediction.

Specifically, the method first obtains historical route data and corresponding historical state results of the target driving vehicle. Then, use machine learning methods to train these data to obtain a route prediction model. The input variable of the route prediction model is the state prediction result, and the output result is the predicted route.

In practical applications, when the real-time vehicle state data of the target driving vehicle is obtained and the vehicle state is predicted, the prediction result can be used as an input variable and input into the preset route prediction model to obtain the predicted route. Then, the predicted route is provided to the remote end for intelligent traffic control, so as to realize more intelligent and efficient traffic control.

In short, this method uses machine learning methods for route prediction, and obtains a route prediction model by training historical route data and corresponding historical state results. In practical applications, the predicted route can be obtained by inputting the state prediction result as an input variable into the preset route prediction model, thereby providing more comprehensive and accurate information support for intelligent traffic control.

When the state prediction result is used as an input variable and input into the preset route prediction model, the following methods can be used for route prediction:

data preparation

First, the historical route data and corresponding historical state results of the target driving vehicle need to be prepared. These data can be regarded as a training data set, which is used to train the route prediction model.

model training

Using the method of machine learning, the historical route data and the corresponding historical state results are trained to obtain the route prediction model.

In specific implementation, various machine learning algorithms, such as decision tree, neural network, support vector machine, etc., can be used to train the route prediction model.

predicted route

When the real-time vehicle state data of the target driving vehicle is obtained and the vehicle state is predicted, the prediction result can be used as an input variable and input into the preset route prediction model to obtain the predicted route.

Provide the predicted route to the remote end for intelligent traffic control in order to achieve more intelligent and efficient traffic control.

In actual implementation, various remote control algorithms can be used, such as intelligent traffic light control, dynamic route planning, etc., to achieve more efficient traffic control based on the predicted route.

Example 9

On the basis of the previous embodiment, the method of inputting the behavior prediction result as an input variable into the preset behavior judgment model to obtain the driving safety index of the target driver and providing it to the remote end for intelligent traffic warning includes: Assuming that the driving behavior sequence corresponding to the behavior prediction result is S=s ₁ , s ₂ ,..., s _n , where s _i represents the driving behavior state of the target driver at time point i, input this sequence into the behavior judgment model, and get Driving safety index S _c , the formula is as follows:

Among them, N is the length of the behavior sequence, M _i is the number of all possible driving behaviors at time point i, p(j|i) is the probability of predicting the jth driving behavior at time point i, weight _j is the The weight of j kinds of driving behaviors; the judgment model first predicts possible driving behaviors and corresponding probabilities for each time point, and then for each time point, calculates the weighted average value of all possible driving behaviors, wherein, the higher the weight Larger means more dangerous, and finally the average value of all time points is used as the driving safety index.

Example 10

On the basis of the previous embodiment, the remote end is a remote intelligent traffic control center.

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by instructions controlling related hardware, and the instructions can be stored in a computer-readable storage medium, and is loaded and executed by the processor.

To this end, an embodiment of the present invention provides a computer-readable storage medium, which stores a plurality of instructions that can be loaded by a processor to execute any of the methods for generating shopping information provided by the embodiments of the present invention. A step of. Wherein, the storage medium may include: a read only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and the like.

Due to the instructions stored in the storage medium, the steps in any method for generating shopping information provided by the embodiments of the present invention can be executed, so any method for generating shopping information provided by the embodiments of the present invention can be realized For the beneficial effects that can be achieved, refer to the previous embodiments for details, and will not be repeated here.

For the specific implementation of the above operations, reference may be made to the foregoing embodiments, and details are not repeated here.

The above is a detailed introduction to a machine learning-based vehicle driving behavior prediction method provided by the embodiment of the present invention. In this paper, a specific example is used to illustrate the principle and implementation of the present invention. The description of the above embodiment is only for To help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation and application range. In summary, the content of this specification should not be construed as a limitation of the invention.

Claims (10)

1. A method for predicting vehicle driving behavior based on machine learning, characterized in that the method comprises: determining a forecast target, the forecast target comprising: a target driver and a target driving vehicle; obtaining the historical driving behavior data of the target driver and the historical vehicle motion state data of the target driving vehicle; based on the historical driving behavior data, a driving behavior chain is constructed; the driving behavior chain is a tree structure, composed of nodes and connecting lines, each node represents a driving behavior, and the driving behavior In the chain, starting from the root node and ending at each end node, it constitutes a complete single driving behavior chain, which represents all the driving behaviors performed by the target driver at the end of a complete vehicle driving; based on historical vehicle state data , build a vehicle state chain; the vehicle state chain is a tree structure, consisting of nodes and connecting lines, each node represents a vehicle state, in the vehicle state chain, starting from the root node and ending at each terminal node, constitutes A complete single vehicle state chain, which represents all the vehicle states experienced by the target driving vehicle at the end of a complete vehicle driving; obtain the real-time driving behavior data of the target driver, based on the driving behavior chain, the driver The driving behavior of the target vehicle is predicted to obtain the behavior prediction result, and the real-time vehicle motion state data of the target driving vehicle is obtained. On the basis of the vehicle state connection, the vehicle state of the target driving vehicle is predicted to obtain the state prediction result; the state prediction result is used as The input variable is input to the preset route prediction model to obtain the predicted route of the target driving vehicle, which is provided to the remote end for intelligent traffic control; the behavior prediction result is input into the preset behavior judgment model as an input variable to obtain the target driving The driving safety index of personnel is provided to the remote end for intelligent traffic warning. 2. The method according to claim 1, wherein the driving behavior data at least includes: seat belt wearing status, mobile phone use times, eating and drinking times and blinking times; the historical driving behavior data is historical target drivers The driving behavior data; the real-time driving behavior data is the real-time driving behavior data of the target driver. 3. The method according to claim 1, wherein the vehicle motion state data at least includes: position, speed, acceleration and steering angle; the historical vehicle motion state data is the vehicle motion state of the historical target driving vehicle data; the real-time vehicle motion state data is real-time vehicle motion state data of the target driving vehicle. 4. The method according to claim 2 or 3, characterized in that, based on historical driving behavior data, the method for constructing a driving behavior chain comprises: considering the driving behavior chain as a sequence set composed of multiple driving behavior sequences; All historical driving behavior data corresponding to each identical moment in the historical driving behavior data are regarded as a driving behavior sequence; each historical driving behavior data included in the driving behavior sequence is regarded as a driving behavior state, and The driving behavior corresponding to a moment is regarded as an observation value of driving behavior; the driving behavior sequence is regarded as a state sequence, in which the state of each driving behavior is represented as a hidden state, and the observed value of each driving behavior is represented as a Observation states; all possible sequences of hidden states are inferred as driving behavior chains using hidden Markov models. 5. The method according to claim 4, characterized in that, the method of obtaining the real-time driving behavior data of the target driver, predicting the driving behavior of the driver on the basis of the driving behavior chain, and obtaining the behavior prediction result Including: in the driving behavior chain, according to the current driving state and transition matrix, calculate the probability distribution of the next driving state; express the current state as a probability vector, and then calculate the probability distribution of the next state through vector multiplication; According to the calculated probability distribution of the next state, select the state with the highest probability as the next driving state of the target driver, find the state with the highest probability in the probability distribution, and then use it as the next state; repeat the above steps until the prediction The period of time ends; in this process, the driving state of the target driver at each time point is recorded, and the behavior sequence is obtained as the prediction result of the target driver's future behavior. 6. The method according to claim 2 or 3, wherein, based on historical vehicle motion state data, the method for constructing a vehicle state chain comprises: converting historical vehicle motion state data into a discretized state; specifically, Convert continuous real values such as vehicle position, velocity, acceleration and steering angle into corresponding discrete states; regard the vehicle state sequence as a hidden Markov model, assuming vehicle state sequence s ₁ , s ₂ ,..., s _T is a first-order Markov chain of length T, where s _t represents the vehicle state at time t; estimate model parameters A, B and π, where A is the state transition matrix, B is the observation probability matrix, and π is the initial state Probability vector; use historical vehicle motion state data to estimate model parameters, so that the likelihood value of the model on the observed state sequence is the largest; use model parameters and real-time vehicle motion state data to predict vehicle state; use forward algorithm to calculate the Under the condition that the observed state at the current moment is o ₁ , o ₂ ,..., o _t , the posterior probability P(s _t =i|o ₁ , o ₂ ,..., o _t , A, B, π), and then determine the vehicle state at the current moment according to the maximum a posteriori probability criterion, namely: Using the predicted vehicle state sequence, the continuous same state is merged into one state node, thus constructing the vehicle state chain. 7. The method according to claim 6, wherein the acquisition of the real-time vehicle motion state data of the target driving vehicle predicts the vehicle state of the target driving vehicle on the basis of the vehicle state connection, and obtains a state prediction result The method includes: in the vehicle state chain, according to the current vehicle state and the transition matrix, calculate the probability distribution of the next vehicle state; express the current state as a probability vector, and then calculate the probability of the next state through vector multiplication distribution; according to the calculated probability distribution of the next state, select the state with the highest probability as the next vehicle state of the target driving vehicle, find the state with the highest probability in the probability distribution, and then use it as the next state; repeat the above steps, Until the end of the predicted time period; in this process, record the vehicle state of the target driving vehicle at each time point, so as to obtain the vehicle state sequence as the prediction result of the future behavior of the target driving vehicle. 8. The method according to claim 7, wherein the state prediction result is used as an input variable and input to a preset route prediction model to obtain a predicted route of the target driving vehicle, which is provided to the remote end for intelligent traffic control The method includes: obtaining historical route data and corresponding historical state results of a target driving vehicle, and using machine learning to train to obtain a route prediction model, the input variable of the route prediction model is a state prediction result, and the output result is a predicted route. 9. The method according to claim 5, wherein the behavior prediction result is input into a preset behavior judgment model as an input variable to obtain the driving safety index of the target driver and provide it to the remote end for intelligent The traffic early warning method includes: assuming that the driving behavior sequence corresponding to the behavior prediction result is S=s ₁ , s ₂ , ..., s _n , where s _i represents the driving behavior state of the target driver at time point i, and the sequence Input it into the behavior judgment model to get the driving safety index S _c , the formula is as follows: Among them, N is the length of the behavior sequence, M _i is the number of all possible driving behaviors at time point i, p(j|i) is the probability of predicting the jth driving behavior at time point i, weight _j is the The weight of j kinds of driving behaviors; the judgment model first predicts possible driving behaviors and corresponding probabilities for each time point, and then for each time point, calculates the weighted average value of all possible driving behaviors, wherein, the higher the weight Larger means more dangerous, and finally the average value of all time points is used as the driving safety index. 10. The method according to claim 8 or 9, wherein the remote end is a remote intelligent traffic control center.