CN109948550A - A system and method for monitoring the flow of people in a smart railway station - Google Patents

Abstract

The invention discloses a system and a method for monitoring the pedestrian flow of an intelligent railway station, which comprises a data acquisition module for acquiring videos of all positions of the railway station, a data analysis platform for carrying out face detection, face quantity statistics, pedestrian flow analysis and prediction on the videos and a monitoring terminal. By the method and the device, the monitoring resources such as the camera and the like can be fully utilized to automatically monitor the human flow, the accuracy of human face detection is higher, and manpower and material resources are saved.

Description

A system and method for monitoring the flow of people in a smart railway station

technical field

The invention relates to the technical field of monitoring, in particular to a system for monitoring the flow of people in a smart railway station.

Background technique

In daily life, people go to various public places every day, ranging from shopping malls, airports and railway stations, to subway stations, or to certain concerts and competition venues. These are places with dense traffic, especially train stations, which are subject to a large number of traffic every day. Managers need to focus on this to prevent unexpected situations such as congestion and stampede.

The traditional method is that the management party arranges a large number of manpower to conduct patrols, effectively divert the flow of people, and prevent the occurrence of emergencies. At the same time, the cameras installed in various locations are used for auxiliary monitoring, and special personnel are arranged to view them in real time. However, relying on manpower to implement management in areas with high concentration of people not only consumes manpower and increases costs, but also is not accurate, and does not make full use of surveillance resources such as cameras. Once an emergency occurs, it is not possible to grasp the specific value of the flow of people in the area in time, and it is difficult to determine what level of evacuation and emergency measures to take.

The current mainstream face detection technology is used to develop and train models using the facial features of yellow and white people, and the lighting conditions in the samples are good, which can better detect yellow and white people from the video. Face. However, the existing algorithms and technologies have flaws in the detection of black faces, unable to cope with black people and poor lighting conditions, and cannot detect black faces in camera videos and photos, especially in complex backgrounds and backlighting. Under such conditions, the detection of black faces is more difficult. For the face detection of black people, as of now, there is no complete set of effective solutions in the industry.

SUMMARY OF THE INVENTION

In view of the deficiencies of the prior art, the present invention aims to provide a system for monitoring the flow of people in a smart railway station, which can make full use of monitoring resources such as cameras to automatically monitor the flow of people, with higher accuracy of face detection and saving manpower and material resources. .

In order to achieve the above object, the present invention adopts the following technical solutions:

A smart railway station people flow monitoring system, comprising:

Data acquisition module: including camera devices deployed in various positions of the railway station and corresponding controllers, the camera devices are used to obtain the video data of the location and transmit it to the controller connected to it, and the controllers are respectively transmitted to the data analysis platform and monitoring terminal;

Data analysis platform: including a face detection model, a big data analysis module, a map component module and a storage module; the face detection model is used to detect all the face data in the video data obtained by each camera device and store it in the In the storage module; the big data analysis module is used to count the number of faces in each video data stored in the storage module, so as to obtain the number of faces in the video data obtained by the camera device at each position at each time point, Then, the data of people flow and the trend of people flow at each location of the railway station at each time point are obtained by statistics, and the flow of people at each location in the next set period of time is predicted according to the trend of people flow in the set time period. When the traffic in a certain location exceeds the preset traffic threshold, an alarm will be issued and the location of the alarm will be quickly indicated in combination with the map component module; the map component module is used to store the location of each camera device;

The face detection model adopts the MTCNN model, which is obtained by training in the following manner: first collect the face data of Caucasians, yellow people, and black people, and then use the face data of Caucasians and yellow people to carry out the MTCNN model. Preliminary training, after reaching the set accuracy requirements, adopt the transfer learning method, and then use the face data of black people to retrain the MTCNN model obtained from the preliminary training to obtain the required face detection model;

Monitoring terminal: It is used to receive the video data collected by the data acquisition module at each location of the railway station at various time points and display it for real-time monitoring, and to obtain the next set period of time from the data analysis platform. Personal traffic trend prediction results and alarm information are displayed.

Further, the data acquisition module further includes a lighting device, and the lighting device and the camera device are installed at various positions of the railway station together and connected to the controller.

Further, the training accuracy of the MTCNN model is improved by the online hard sample mining method, specifically: in the training process, the errors in each training batch are sorted in descending order, and the first 70% of the sample errors in the sorting result are taken out. , carry out error transfer, and correct the parameters of the MTCNN model.

Further, in the training of the MTCNN model, the input image is scaled.

Further, in the training of the MTCNN model, the method of focusing on difficult samples is adopted.

The present invention also provides a method for monitoring the flow of people in a railway station by using the above system, comprising the following steps:

In step S1, the camera devices deployed in various positions of the railway station acquire the video data of the location and transmit it to the connected controller, the controller transmits the video data to the data analysis platform and the monitoring terminal, and the monitoring terminal displays the video data for the staff. real time monitoring;

In step S2, in the data analysis platform, the face detection model stores all the face data in the video data acquired by the detected camera devices in the storage module; the big data analysis module stores all the face data in the storage module. The number of faces in the video data is counted, so as to obtain the number of faces in the video data obtained by the camera device at each location at each time point, and then the statistics of the traffic data and the traffic trend of each location of the railway station at each time point are obtained. , and predict the flow of people at each location in the next set time period according to the trend of the flow of people in the set time period. When the predicted flow of people in a certain location will exceed the preset threshold of people flow, an alarm will be issued and combined with The map component module quickly indicates the location of the alarm; the data analysis platform will transmit the prediction results of the traffic trend and the alarm information to the monitoring terminal in real time for display for the staff to view.

The beneficial effects of the present invention are:

1. The system of the present invention performs real-time processing on the video data transmitted back by the camera device installed in each area, and uses a specially designed training method to train the face detection model to detect all the faces contained in the video. Through big data analysis Count, count and analyze the faces detected by the model, depict the flow trend of people in the area, and call historical data for back-checking at any time. When it is predicted that the flow of people in a certain area reaches the preset warning value, it will alarm in time, and after combining with the regional map, the accurate location and arrival route will be given, and personnel will be arranged to handle it accurately to realize intelligent control.

2. The face detection model of the system of the present invention can accurately detect yellow people, white people and black people appearing in the video, and accurately detect all the faces contained in the video, so that the flow of people in the monitored area can be realized. Perform accurate statistics and analysis.

Description of drawings

1 is a schematic diagram of the system composition of Embodiment 1 of the present invention;

Fig. 2 is the schematic flow chart of the method in Embodiment 2 of the present invention;

3 is a schematic diagram of a training flow of the MTCNN model in Embodiment 1 of the present invention;

Figure 4 is a schematic diagram of the overall architecture of the MTCNN model.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the present embodiment takes the technical solution as the premise, and provides a detailed implementation manner and a specific operation process, but the protection scope of the present invention is not limited to the present invention. Example.

The following will briefly explain the technical terms involved in the embodiment:

Transfer learning refers to taking a trained model, making some directional adjustments, and then transferring it to a related problem.

The latitude refers to the range of the brightness contrast of the scene that the film can correctly accommodate. The higher the latitude, the greater the brightness contrast can be accommodated.

Example 1

This embodiment provides a system for monitoring the flow of people in a smart railway station, as shown in FIG. 1 , including:

Data acquisition module: including camera devices deployed in various positions of the railway station and corresponding controllers, the camera devices are used to obtain the video data of the location and transmit it to the controller connected to it, and the controllers are respectively transmitted to the data analysis platform and monitoring terminal;

Data analysis platform: including a face detection model, a big data analysis module, a map component module and a storage module; the face detection model is used to detect all the face data in the video data obtained by each camera device and store it in the In the storage module; the big data analysis module is used to count the number of faces in each video data stored in the storage module, so as to obtain the number of faces in the video data obtained by the camera device at each position at each time point, Then, the data of people flow and the trend of people flow at each location of the railway station at each time point are obtained by statistics, and the flow of people at each location in the next set period of time is predicted according to the trend of people flow in the set time period. When the traffic in a certain location exceeds the preset traffic threshold, an alarm will be issued and the location of the alarm will be quickly indicated in combination with the map component module; the map component module is used to store the location of each camera device;

The face detection model adopts the MTCNN model, which is obtained by training in the following way: because the currently publicly accessible face data sets are mostly yellow and white, and black face data is very scarce. Transfer learning can effectively deal with the problem of uneven data and data scarcity. First collect the face data of white, yellow and black people, and then use the face data of white and yellow people to perform preliminary training on the MTCNN model. After reaching the set accuracy requirements, the transfer learning method is adopted. , and then use the face data of black people to retrain the MTCNN model obtained from the initial training to obtain the required face detection model, as shown in Figure 3.

Specifically, when the MTCNN model obtained by preliminary training is transferred to the face data of black people for retraining, the model parameters can be fine-tuned according to actual needs.

Monitoring terminal: It is used to receive the video data collected by the data acquisition module at each location of the railway station at various time points and display it for real-time monitoring, and to obtain the next set period of time from the data analysis platform. Personal traffic trend prediction results and alarm information are displayed.

In this embodiment, the data acquisition module further includes a lighting device, and the lighting device and the camera device are installed at various positions of the railway station together and connected to the controller. By setting the lighting device, it is possible to change the lighting conditions to make the black people easier to capture in the camera picture, thereby improving the performance of the black people's face detection. It is also possible to further adopt a high-tolerance camera device, which can make the faces of black people more obvious in the video, and the detection is more accurate.

In this embodiment, the training accuracy of the MTCNN model is also improved through the online hard sample mining method, specifically: in the training process, the errors in each training batch are sorted in descending order, and the top 70% of the sorting results are sorted in descending order. The sample error is taken out, the error is transmitted, and the parameters of the MTCNN model are corrected.

In this embodiment, in the training of the MTCNN model, the input image is scaled, so that the trained face detection model has better accuracy under different image scales.

In this embodiment, in the training of the MTCNN model, the method of focusing on difficult samples is adopted, that is, when the face recognition error is not large (it can be determined by being less than a set threshold), parameter correction is not performed, Parameter correction is performed for face samples with large errors (which can be determined by greater than or equal to a set threshold). Due to the darker skin color of black people, many people overlap each other, and there is a situation of false detection or missed detection. Using the above method can make the trained model have better recognition ability for difficult samples.

It should be noted that the overall network structure of the MTCNN model is shown in Figure 4. The original image needs to go through three network layers, and the networks adopt a cascade structure, that is, multiple convolutional network models (CNN) are connected in series. Perform multiple predictions on the input image to continuously improve the accuracy of face detection tasks. The MTCNN model is divided into three parts in structure, namely P-Net, R-Net, and O-Net. Each part has three tasks, namely predicting whether it is a face or not, and predicting the positioning point of the face bounding box. And the five key points of the face (left eye, right eye, nose tip, left mouth corner, right mouth corner) are predicted, which can improve the accuracy of face detection from different angles. The three parts will filter out some non-face parts each time, and make further corrections based on the key points of the original prediction, so that the model has both real-time and accuracy.

Each level of the MTCNN model has corresponding classification information, bounding box information and face key point information, each of which has its own loss function.

(1) For classification error:

This function is an error function for classification, called log loss function or log likelihood loss function, also known as logistic regression loss function or cross entropy loss function. The function is defined on a probability estimate, with a logarithmic function transforming the value into a confidence probability for the class. Logarithmic loss quantifies the accuracy of a classifier by penalizing incorrect classifications. Minimizing the log loss is basically equivalent to maximizing the accuracy of the classifier. When calculating the loss, the input is the probability value for each class to which it belongs.

Perform a binary classification judgment on whether the detection area contains a face. If it contains a face and the model also predicts that it contains a face, the error is 0. Otherwise, the error is not 0. By continuously optimizing the error function, it is possible to determine whether it contains a face. Precise classification.

(2) For bounding box regression error:

This function is the mean square error function, which is often used in regression problems. It is the expected value of the square of the difference between the estimated value of the parameter and the true value of the parameter. This function is defined based on the distance, and judges whether the predicted value is accurate by the square error of the corresponding element between the predicted value and the actual value, and realizes the quantification of the regressor. Optimizing the mean squared error function can reduce the error of the predicted value, and when calculating the loss, the input is the predicted value.

For the regression of the upper left and lower right coordinate points of the bounding box of the face, the difference between the corresponding coordinate points is the Euclidean distance. The smaller the difference, the more accurate the regression and the smaller the error, and vice versa.

(3) For the error of the key points of the face:

This function is similar to the error function of the bounding box, which is also the mean square error function. By optimizing this function, the regression error of the key points of the face is minimized. The optimization method for face key points is similar to the optimization method for bounding boxes. If the prediction of key points is more accurate, the error will be smaller, and vice versa.

By optimizing these three loss functions, the overall error is minimized to achieve the effect of face detection. The algorithm can detect the face in the input image, and return the corresponding face bounding box and face key points to complete the face detection task.

Example 2

This embodiment provides a method for monitoring the flow of people in a railway station by using the system described in Embodiment 1, as shown in FIG. 2 , including the following steps:

In step S1, the camera devices deployed in various positions of the railway station acquire the video data of the location and transmit it to the connected controller, the controller transmits the video data to the data analysis platform and the monitoring terminal, and the monitoring terminal displays the video data for the staff. real time monitoring;

In step S2, in the data analysis platform, the face detection model stores all the face data in the video data acquired by the detected camera devices in the storage module; the big data analysis module stores all the face data in the storage module. The number of faces in the video data is counted, so as to obtain the number of faces in the video data obtained by the camera device at each location at each time point, and then the statistics of the traffic data and the traffic trend of each location of the railway station at each time point are obtained. , and predict the flow of people at each location in the next set time period according to the trend of the flow of people in the set time period. When the predicted flow of people in a certain location will exceed the preset threshold of people flow, an alarm will be issued and combined with The map component module quickly indicates the location of the alarm; the data analysis platform will transmit the prediction results of the traffic trend and the alarm information to the monitoring terminal in real time for display for the staff to view. The staff can also guide and command the police officers accordingly.

For those skilled in the art, various corresponding changes and deformations can be given according to the above technical solutions and concepts, and all these changes and deformations should be included within the protection scope of the claims of the present invention.

Claims (6)

1. a smart railway station people flow monitoring system, is characterized in that, comprises: Data acquisition module: including camera devices deployed in various positions of the railway station and corresponding controllers, the camera devices are used to obtain the video data of the location and transmit it to the controller connected to it, and the controllers are respectively transmitted to the data analysis platform and monitoring terminal; Data analysis platform: including a face detection model, a big data analysis module, a map component module and a storage module; the face detection model is used to detect all the face data in the video data obtained by each camera device and store it in the In the storage module; the big data analysis module is used to count the number of faces in each video data stored in the storage module, so as to obtain the number of faces in the video data obtained by the camera device at each position at each time point, Then, the data of people flow and the trend of people flow at each location of the railway station at each time point are obtained by statistics, and the flow of people at each location in the next set period of time is predicted according to the trend of people flow in the set time period. When the traffic in a certain location exceeds the preset traffic threshold, an alarm will be issued and the location of the alarm will be quickly indicated in combination with the map component module; the map component module is used to store the location of each camera device; The face detection model adopts the MTCNN model, which is obtained by training in the following manner: first collect the face data of Caucasians, yellow people, and black people, and then use the face data of Caucasians and yellow people to carry out the MTCNN model. Preliminary training, after reaching the set accuracy requirements, adopt the transfer learning method, and then use the face data of black people to retrain the MTCNN model obtained from the preliminary training to obtain the required face detection model; Monitoring terminal: It is used to receive the video data collected by the data acquisition module at each location of the railway station at various time points and display it for real-time monitoring, and to obtain the next set period of time from the data analysis platform. Personal traffic trend prediction results and alarm information are displayed. 2 . The people flow monitoring system of a smart railway station according to claim 1 , wherein the data acquisition module further comprises a lighting device, and the lighting device and the camera device are installed at various positions of the railway station together. 3 . , and connected to the controller. 3. The smart railway station people flow monitoring system according to claim 1 is characterized in that, the training accuracy of the MTCNN model is also improved by the onlinehard sample mining method, specifically: in the training process, the The errors are sorted in descending order, and the first 70% of the sample errors in the sorting result are taken out, and the errors are transferred to correct the parameters of the MTCNN model. 4 . The system for monitoring the flow of people in a smart railway station according to claim 1 , wherein in the training of the MTCNN model, the input image is scaled. 5 . 5 . The system for monitoring the flow of people in a smart railway station according to claim 1 , wherein, in the training of the MTCNN model, a method of focusing on difficult samples is adopted. 6 . 6. a method utilizing the system described in any of the preceding claims to carry out the monitoring of people flow in railway station, is characterized in that, comprises the steps: In step S1, the camera devices deployed in various positions of the railway station acquire the video data of the location and transmit it to the connected controller, the controller transmits the video data to the data analysis platform and the monitoring terminal, and the monitoring terminal displays the video data for the staff. real time monitoring; In step S2, in the data analysis platform, the face detection model stores all the face data in the video data acquired by the detected camera devices in the storage module; the big data analysis module stores all the face data in the storage module. The number of faces in the video data is counted, so as to obtain the number of faces in the video data obtained by the camera device at each location at each time point, and then the statistics of the traffic data and the traffic trend of each location of the railway station at each time point are obtained. , and predict the flow of people at each location in the next set time period according to the trend of the flow of people in the set time period. When the predicted flow of people in a certain location will exceed the preset threshold of people flow, an alarm will be issued and combined with The map component module quickly indicates the location of the alarm; the data analysis platform will transmit the prediction results of the traffic trend and the alarm information to the monitoring terminal in real time for display for the staff to view.