CN116153086B - Multi-section traffic accident and congestion detection method and system based on deep learning - Google Patents

Abstract

The invention relates to the technical field of traffic control, and provides a method and a system for detecting traffic accidents and congestion of multiple sections based on deep learning, wherein the detection method comprises the following steps: sequentially accessing the cameras by adopting a polling mechanism and congestion pre-judgment, and acquiring vehicle images acquired by the cameras; according to the acquired vehicle image, a trained traffic event three-classification model is adopted to carry out vehicle detection and identification, and a traffic jam condition is obtained; acquiring a vehicle image of a traffic jam road section, identifying by adopting a second classification model of the jam accident direction to obtain a lane middle line or a separation zone and a vehicle head direction, and determining the jam accident direction; and the traffic event three-classification model sequentially identifies vehicles in the image, the number of the vehicles in the image and the moving distance of the vehicles, and fuses the number of the vehicles and the moving distance of the vehicles to obtain traffic jam conditions. The detection method can process mass traffic monitoring data by adopting a small amount of server resources, and can accurately control the congestion direction of the road.

Description

Multi-section traffic accident and congestion detection method and system based on deep learning

technical field

The present invention relates to the technical field related to traffic control, in particular to a deep learning-based multi-section traffic accident and congestion detection method and system.

Background technique

With the continuous increase of people's travel demand and the construction of more and more expressways, the workload and difficulty of traditional manual methods to deal with expressway traffic congestion and accident identification are also gradually increasing. Traditional methods mainly answer by telephone or manually check Detection is done by way of surveillance video, and the workload and difficulty of manual viewing are huge, which consumes a lot of labor costs. Even if some current solutions use artificial intelligence automatic detection methods, the recognition efficiency will be relatively low due to the limitations of computer performance and computer hardware. To improve the recognition efficiency, it is necessary to invest more GPU servers to achieve it, which increases the The cost of construction, in the face of a large number of surveillance cameras, detection will be relatively slow, and real-time performance cannot be guaranteed.

Contents of the invention

In order to solve the above problems, the present invention proposes a multi-section traffic accident and congestion detection method and system based on deep learning, which can process massive traffic monitoring data with a small amount of server resources, and can accurately identify the direction of road congestion.

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

One or more embodiments provide the deep learning-based multi-section traffic accident and congestion detection method, comprising the following steps:

Use the polling mechanism and congestion pre-judgment to access the camera in turn to obtain the vehicle images collected by the camera;

According to the acquired vehicle images, the trained traffic event three-classification model is used for vehicle detection and recognition, and the traffic congestion situation is obtained;

Obtain the vehicle image of the traffic jam section, use the congestion accident direction binary classification model to identify, identify the middle line of the lane or the isolation zone, and the direction of the front of the vehicle, and determine the congestion accident direction;

The three-classification model of traffic events sequentially identifies the vehicles in the image, the number of vehicles in the image, and the moving distance of the vehicles, and combines the number of vehicles and the moving distance of the vehicles to obtain the traffic congestion situation.

One or more embodiments provide a multi-section traffic accident and congestion detection system based on deep learning, including:

Camera polling control module: configured to use polling mechanism and congestion pre-judgment to sequentially access the camera to obtain vehicle images collected by the camera;

Congestion situation identification module: configured to use the trained traffic event three-classification model for vehicle detection and recognition based on the acquired vehicle images to obtain traffic congestion conditions;

Congestion accident direction recognition module: configured to obtain vehicle images of traffic congestion road sections, use the congestion accident direction binary classification model to identify, identify the middle line of the lane or the isolation zone, and the direction of the front of the vehicle, and determine the congestion accident direction;

The three-classification model of traffic events sequentially identifies the vehicles in the image, the number of vehicles in the image, and the moving distance of the vehicles, and combines the number of vehicles and the moving distance of the vehicles to obtain the traffic congestion situation.

An electronic device includes a memory, a processor, and computer instructions stored in the memory and run on the processor. When the computer instructions are executed by the processor, the steps described in the above methods are completed.

A computer-readable storage medium is used for storing computer instructions, and when the computer instructions are executed by a processor, the steps described in the above method are completed.

Compared with prior art, the beneficial effect of the present invention is:

In the present invention, the camera polling mechanism is innovatively used to quickly identify congestion and accident events in massive monitoring data, and only a small number of GPU servers are needed to realize real-time detection of all cameras on the high-speed road section, which can ensure the efficiency of detection , can guarantee the accuracy of detection, and solve the problem that a large amount of server resources are required to use the congestion accident detection service. In addition, the two conditions of incoming vehicle direction, outgoing vehicle direction and the left and right sides of the isolation belt are innovatively used to accurately determine the road where congestion or accidents occur in the image.

Advantages of the present invention, as well as advantages of additional aspects, will be described in detail in the following specific examples.

Description of drawings

The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention, and the schematic embodiments of the present invention and their descriptions are used to explain the present invention, but not to limit the present invention.

Fig. 1 is the flow chart of the multi-section traffic accident and congestion detection method of embodiment 1 of the present invention;

2 is a schematic diagram of a multi-section traffic accident and congestion detection process in Embodiment 1 of the present invention;

Fig. 3 is a road schematic diagram of an example of congestion direction recognition in Embodiment 1 of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the present invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It should be noted that the terminology used here is only for describing specific embodiments, and is not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural, and it should also be understood that when the terms "comprising" and/or "comprising" are used in this specification, they mean There are features, steps, operations, means, components and/or combinations thereof. It should be noted that, in the case of no conflict, various embodiments and features in the embodiments of the present invention can be combined with each other. The embodiments will be described in detail below in conjunction with the accompanying drawings.

Example 1

In the technical solution disclosed in one or more embodiments, as shown in Figures 1 to 3, the multi-section traffic accident and congestion detection method based on deep learning includes the following steps:

Step 1. Use the polling mechanism and congestion pre-judgment to access the camera in turn to obtain the vehicle images collected by the camera;

Step 2. According to the acquired vehicle image, use the trained traffic event three-classification model to detect and identify the vehicle to obtain the traffic congestion situation;

Step 3. Obtain the vehicle image of the traffic jam road section, and use the congestion accident direction binary classification model to identify, identify the middle line of the lane or the isolation zone, and the direction of the front of the vehicle, and determine the congestion accident direction.

The vehicle detection by the three-classification model of traffic events includes sequentially identifying the vehicles in the image, the number of vehicles in the image, and the moving distance of the vehicles, and fusing the number of vehicles and the moving distance of the vehicles to obtain traffic congestion.

In this embodiment, the camera polling mechanism is innovatively used to quickly identify congestion and accident events in a large amount of monitoring data. Only a small number of GPU servers are needed to realize real-time detection of all cameras on the high-speed road section, which can ensure the efficiency of detection. It can guarantee the accuracy of detection and solve the problem that a large amount of server resources are required to use the congestion accident detection service. Among them, using 2 to 4 GPU servers, the number of detection cameras can reach the order of thousands. In addition, the innovative use of conditions such as the direction of incoming traffic, the direction of outgoing traffic, and the left and right sides of the barrier can accurately determine which road in the image is congested or has an accident.

In step 1, the polling mechanism is used to access the cameras sequentially, and the dichotomy method is used to predict the possibility of road congestion in each frame of image. Predict that the camera area of the current camera is congested, and continue to acquire the vehicle image data of the current camera. That is: poll the access camera first, judge whether the image area collected by the current camera is congested, switch to the next camera poll if not congested, and continue to collect the current camera image if congested.

Specifically, the possibility of road congestion in each frame of image is predicted. The prediction method is to identify the number of vehicles in the picture, set the value of the smoothness confidence coefficient according to the number of vehicles, and judge whether it is smooth according to the value of the smoothness confidence coefficient.

Whether the road is clear or not is marked by the smooth confidence coefficient σ1. Due to the use of dichotomy, the confidence coefficient marked as congestion is σ2, and follows σ1+σ2=1.

A specific implementation method, the calculation formula of the unimpeded confidence coefficient σ1 can be as follows:

0<sum(number of vehicles)≤A: σ1=0.9;

A<sum (number of vehicles)≤B: σ1=0.7;

B<sum (number of vehicles)≤C: σ1=0.4;

C<sum(number of vehicles): σ1=0.2;

Among them, A<B<C is the set value.

A technical solution that can be realized, if the number of vehicles is less than or equal to 5, the unimpeded confidence coefficient is marked as 0.9; if the number of vehicles is greater than 5 and less than or equal to 15, the unimpeded confidence coefficient is marked as 0.7; It is marked as 0.4, the number of vehicles is greater than 20, and the unimpeded confidence coefficient is marked as 0.2.

In some embodiments, a first threshold and a second threshold of the unimpeded confidence coefficient are set, and the first threshold is greater than the second threshold;

If the calculated unobstructed confidence coefficient is not less than the set first threshold, the road in the camera area of the camera is directly marked as unimpeded, and the next camera image is directly switched for analysis; the unimpeded confidence coefficient is between the first threshold and the second threshold During the period, continue to obtain the first set frame number image data of the current camera to calculate the average value of the unimpeded confidence coefficient, determine whether the current road is congested according to the average value of the unimpeded confidence coefficient, and switch to the next camera for image acquisition; the unimpeded confidence coefficient is not Greater than the second threshold, continue to obtain the second set frame number image data of the current camera to calculate the average value of the unimpeded confidence coefficient, determine whether the current road is congested according to the average value of the unimpeded confidence coefficient, and switch to the next camera for image acquisition;

Wherein, the first set frame number is greater than the second set frame number.

Specifically, in this embodiment, the first threshold of the unimpeded confidence coefficient may be set to 0.9, and the second threshold may be set to 0.5.

If the current camera is unblocked and the confidence coefficient σ1≥0.9, the road of this camera is directly marked as unblocked, and the image analysis of the next camera is directly switched.

If the camera’s unblocked confidence coefficient is 0.9>σ1>0.5, then the unobstructed confidence coefficient obtained from the image analysis is not high. The smooth confidence coefficients of the pictures are summed and averaged, and the average is marked as , if /> , it is defined as unimpeded, if , it is defined as congestion.

If the camera’s unimpeded confidence coefficient σ1≤0.5, the unimpeded confidence coefficient obtained from the image analysis is extremely low. Confidence coefficient, the smooth confidence coefficient of these 10 pictures is summed and averaged, and the average is marked as , if /> , it is defined as unimpeded, if /> , it is defined as congestion.

In this embodiment, two intervals are set for the unimpeded self-confidence coefficient, wherein the interval not less than the set first threshold of the unimpeded self-confidence coefficient is the interval determined to be unobstructed; the unimpeded self-confidence coefficient is between the first threshold and the second threshold The interval of is the interval that may be unblocked, and this interval should be monitored in order to achieve the accuracy of monitoring.

In step 2, the three-classification model of traffic events is used to detect and recognize the acquired vehicle images to obtain traffic congestion.

Specifically, the three-classification model of traffic events can use the Caffe general deep learning framework to classify traffic flow, congestion, and accidents;

The three-classification model of traffic events includes sequentially connected vehicle identification network, vehicle counting module, vehicle movement identification module and fusion output module;

The vehicle recognition network is used to recognize the vehicle in the vehicle image, and use the target frame to select the vehicle target;

The vehicle counting module is used to count the vehicle target frames selected by the vehicle recognition network frame;

The vehicle movement recognition module is used to identify the movement distance of the same vehicle in adjacent frame images according to the vehicle target frame selected by the vehicle recognition network;

The fusion output module is used for judging the congestion situation according to the vehicle displacement, judging whether an accident occurs, or/and identifying the congestion mileage.

In some embodiments, the vehicle identification network realizes the identification and detection of various types of vehicles through detection algorithms, and establishes a vehicle identification and detection model for identifying the number and location of vehicles. The vehicle recognition network adopts the Caffe general deep learning framework, and the training process of the vehicle recognition network is as follows:

21) Obtain the video data collected by the camera, and parse it into image data as a data set through the video codec module;

Among them, the dataset contains different types of vehicles.

22) Mark various types of vehicles in the images of the dataset;

Specifically, 10,000 images may be selected to mark various types of vehicles in the images.

23) Extract the characteristics of the vehicle based on the labeled data, and regularize the characteristics of the vehicle;

24) For the data after feature regularization, divide the training set and verification set of the vehicle recognition network;

25) Use the Caffe general deep learning framework to build a vehicle recognition network;

26) Train the vehicle recognition network, identify the vehicle information in each image, and obtain the network parameters after training;

27) Test the trained vehicle recognition network based on the images of the validation set until the accuracy requirement is met.

The vehicle counting module is configured to count the vehicles in the image. Specifically, for the recognition results output by the vehicle recognition network, a counter sum is set up, and the initial value of the counter is set to 0. Every time a vehicle is detected, the value of the counter is increased by 1 , that is, sum=sum+1; after the recognition is completed, the sum value is the number of vehicles in the image. The number of vehicles output by this module can be used for the calculation of the smooth flow confidence coefficient.

The vehicle movement identification module is used to identify whether the vehicle is moving, and can identify the moving distance of the same vehicle in two adjacent frames of images; specifically, the identification method can be as follows:

2.1) For the current camera, according to the set time interval, the adjacent frames of the video are obtained, and the video image data is extracted;

Optionally, the set time interval can be set to several seconds, preferably, it can be set to 1 second.

2.2) Detect two adjacent frames of images, the pixel displacement of the rectangle with the same size as the vehicle target frame is the vehicle’s movement displacement value, and the movement speed can be obtained according to the interval time;

In some embodiments, the fusion output module realizes the identification of congestion and accidents, including the following steps:

211) Set the critical value x of the moving distance of the vehicle;

If the displacement is less than the set critical value x, it can be considered that the vehicle is moving slowly, and if the displacement is zero, it is stationary;

212) Calculate the displacement value of each vehicle, if the vehicle exceeding the set ratio is less than the critical value x, it is judged as congestion;

Optionally, the setting ratio can be 70%, and if the moving distance of more than 70% of the vehicles is less than x, it is congestion.

213) According to the time series, detect the vehicle target frame of the vehicle position, if the set number of vehicles stops and does not move, it is determined that an accident may occur; if the set number of vehicles stops and is identified as congestion, it is determined that an accident has occurred;

Further, the congestion mileage recognition is specifically measured by the distance of the congestion cameras, and the congestion length is calculated by the distance between the cameras. Specifically, the steps are as follows:

21.1) For the camera that detects congestion, obtain the coordinate information of the camera;

21.2) Traverse the cameras adjacent to the congested camera to get all consecutive congested cameras;

21.3) For cameras that are congested continuously, calculate the distance between two cameras that are congested, and the sum of the distances is the congestion mileage.

A congestion camera is a camera that detects congestion. According to the setting rules of the camera coordinates, the distance between the cameras is calculated according to the coordinate information. The cameras have registered coordinate information in the highway management system. For example, the number of a certain camera is G35 TV99 K207+340, where K207+340 represents location information.

Calculate the distance between the two cameras based on the location information of the two cameras. For example, if another camera number is G35 TV99 K209+342, the distance between the two cameras is (209-207)*1K+(342-340) , for 2002 m.

Optionally, mark the camera that detects congestion as BUSYCAM, and traverse its adjacent cameras forward and backward. If the adjacent camera is also congested, continue traversing until the next camera does not detect congestion; mark the group of traversals For the congested cameras, record the locations of all the cameras in the group, obtain the maximum and minimum values of the positions, and the difference between the maximum and minimum values is the distance of the congested cameras, that is, the length of the congestion.

Traditional AI congestion detection can only identify the congestion in the screen, but cannot judge the direction of the congestion. Congestion on any side of the isolation zone will be identified as congestion on the road section where the camera is located, but it cannot be determined which side of the isolation zone. That is, which specific road is congested. In addition, the existing monitoring camera is a ball camera, and the camera will rotate at any time, which increases the difficulty of detection.

In step 3, the two-classification model of the direction of the congestion accident is used to classify whether the left road or the right road of the congested section is congested. Specifically, it is configured to perform the following process:

31) Use Faster-RCNN to identify the isolation zone or the middle line;

Faster-RCNN is a target detection algorithm. On the basis of Fast-RCNN, an RPN candidate frame generation algorithm is proposed, which greatly improves the target detection speed.

32) Segment the image with the identified isolation zone or middle line to obtain two new images;

Optionally, OpenCV can be used to segment the image from the isolation zone;

OpenCV (Open source Computer Vision Library, open source computer vision library) is a set of open source API function libraries about computer vision.

33) For each segmented image, calculate the smoothness confidence coefficient according to the number of vehicles in the image, and judge whether the road corresponding to the image is congested;

34) Identify the direction of the vehicle head in the image and determine the direction of the congested road.

The calculation method of the unimpeded confidence coefficient is the same as the calculation method in step 1, and will not be repeated here.

Optionally, it is possible to judge whether the road is smooth by setting a threshold, and judge whether it is smooth according to the value of the smoothness confidence coefficient. The judgment method is the same as above. For a specific example, if σ1≥0.9, the area corresponding to the image is unblocked. If the confidence coefficient of any side is 0.5<σ1<0.9, continue to continuously take 20 frames of images from the camera for further analysis, and then average them as above value, find the average , if /> , it is defined as unimpeded, if /> , it is defined as congestion. If only the left confidence factor /> , it is judged as congestion on the left side, if only the confidence coefficient on the right side /> , it is judged as congestion on the right side. If both left and right confidence coefficients are , it is judged as two-way congestion.

The two-category model of congestion accident direction includes a middle line or isolation zone identification module, a vehicle front identification module and an accident direction determination module.

This embodiment innovatively adopts the middle line or the middle isolation belt and the way of detecting the direction of the front of the vehicle to detect which of the two monitored roads is congested.

The intermediate line or median line identification module is specifically configured to use the Faster-RCNN algorithm to identify and extract the intermediate isolation zone or intermediate line; to identify whether the identified intermediate isolation zone or intermediate line is congested on the left side of the road or on the right side congestion.

Taking the isolation belt as an example, the identification method of the high-speed isolation belt adopts the image feature recognition method. The road on the expressway is gray and black, the middle isolation belt is green, and the middle guardrail is mostly silver or green. There are obvious color boundaries between the two sides and the middle. , the isolation zone can be identified by using the Faster-RCNN algorithm according to the color change.

Vehicle front recognition module: it is configured to use the Faster-RCNN algorithm to identify the direction of the vehicle head according to the images of adjacent frames collected by the camera, and then identify the direction of the incoming vehicle or the direction of the outgoing vehicle.

Optionally, the method for judging the orientation of the vehicle head can be as follows: in chronological order, if the outline of the same vehicle in two adjacent frames of images is getting larger and larger, then the car in the image is the front of the vehicle, and if the outline is getting smaller and smaller, then it is rear.

The accident direction judging module is configured to determine the direction of the congestion accident according to the current orientation of the camera, whether the identified lane is congested on the left or right, and the direction of the front of the vehicle.

The specific recognition example is shown in Figure 3. If there is congestion on the left side of the isolation strip, and it is in the direction of incoming traffic, it is congested on road 1. If the camera rotates 180°, the camera recognizes that it is congested on the right side, and it is in the direction of incoming traffic. When it is judged, the road 1 is congested.

A further technical solution also includes obtaining the staff's feedback data on the traffic accident or congestion data given by the system, and correcting the deviation of accident or congestion judgment. Specifically, build a biased data set and continue to strengthen the training model to enhance the accuracy of accident or congestion recognition. In the initial stage, human participation is performed, and verification is performed based on machine scoring results; congestion and accident probabilities are manually adjusted; and the model is re-corrected based on the adjusted scores to improve the accident prediction effect of the model.

This embodiment realizes the classification of normal driving, accidents, and congestion on the road through the classification algorithm, and can realize the identification of the congestion mileage and the identification of the congestion direction, and can realize more detailed road condition identification. At the same time, the polling mechanism is used to achieve efficient detection. , which improves the real-time performance of detection.

Example 2

Based on embodiment 1, a multi-section traffic accident and congestion detection system based on deep learning is provided in this embodiment, including:

Camera polling control module: configured to use polling mechanism and congestion pre-judgment to sequentially access the camera to obtain vehicle images collected by the camera;

Congestion situation identification module: configured to use the trained traffic event three-classification model for vehicle detection and recognition based on the acquired vehicle images to obtain traffic congestion conditions;

Congestion accident direction recognition module: configured to obtain vehicle images of traffic congestion road sections, use the congestion accident direction binary classification model to identify, identify the middle line of the lane or the isolation zone, and the direction of the front of the vehicle, and determine the congestion accident direction;

The three-classification model of traffic events sequentially identifies the vehicles in the image, the number of vehicles in the image, and the moving distance of the vehicles, and combines the number of vehicles and the moving distance of the vehicles to obtain the traffic congestion situation.

What needs to be explained here is that each module in this embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process is the same, which will not be repeated here.

Example 3

This embodiment provides an electronic device, including a memory, a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps described in the method in Embodiment 1 are completed.

Example 4

This embodiment provides a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the steps described in the method in Embodiment 1 are completed.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, those skilled in the art do not need to pay creative work Various modifications or variations that can be made are still within the protection scope of the present invention.

Claims (8)

1. The multi-section traffic accident and congestion detection method based on deep learning, is characterized in that, comprises the steps: Use the polling mechanism and congestion pre-judgment to access the camera in turn to obtain the vehicle images collected by the camera; According to the acquired vehicle images, the trained traffic event three-classification model is used for vehicle detection and recognition, and the traffic congestion situation is obtained; Obtain the vehicle image of the traffic jam section, use the congestion accident direction binary classification model to identify, identify the middle line of the lane or the isolation zone, and the direction of the front of the vehicle, and determine the congestion accident direction; The three-category model of traffic events sequentially recognizes the vehicles in the image, the number of vehicles in the image, and the moving distance of the vehicles, and combines the number of vehicles and the moving distance of the vehicles to obtain traffic congestion; The polling mechanism is used to access the cameras sequentially. The dichotomy method is used to predict the possibility of road congestion in each frame of image. The road in the camera's camera area is congested, and continue to acquire the vehicle image data of the current camera; Predict the possibility of smooth roads in each frame of image. The prediction method is to identify the number of vehicles in the picture, set the value of the smoothness confidence coefficient according to the number of vehicles, and judge whether it is smooth according to the value of the smoothness confidence coefficient; Setting the first threshold and the second threshold of the unimpeded confidence coefficient, the first threshold is greater than the second threshold; If the smooth confidence coefficient is not less than the set first threshold, the road in the camera area of the camera is smooth, and the next camera screen is directly switched for analysis; The unimpeded confidence coefficient is between the first threshold and the second threshold, continue to obtain the first set frame number image data of the current camera to calculate the average unimpeded confidence coefficient, determine whether the current road is congested according to the average unimpeded confidence coefficient, and switch Go to the next camera for image acquisition; If the unimpeded confidence coefficient is less than the second threshold, continue to obtain the second set frame number image data of the current camera to calculate the average unimpeded confidence coefficient, determine whether the current road is congested according to the average unimpeded confidence coefficient, and switch to the next camera for image processing collection; Wherein, the first set frame number is greater than the second set frame number. 2. the multi-section traffic accident and congestion detection method based on deep learning as claimed in claim 1, is characterized in that: the traffic event three-category model comprises sequentially connected vehicle identification network, vehicle counting module, vehicle movement identification module and fusion output module; The vehicle recognition network is used to recognize the vehicle in the vehicle image, and use the target frame to select the vehicle target; The vehicle counting module is used to count the vehicle target frames selected by the vehicle recognition network frame; The vehicle movement recognition module is used to identify the movement distance of the same vehicle in adjacent frame images according to the vehicle target frame selected by the vehicle recognition network; The fusion output module is used for judging the congestion situation according to the vehicle counting result and the vehicle movement displacement, judging whether an accident occurs, or/and identifying the congestion mileage. 3. the multi-section traffic accident and congestion detection method based on deep learning as claimed in claim 2, is characterized in that, fusion output module realizes congestion and the recognition of accident, comprises the steps: Set the threshold value of vehicle movement distance; Calculate the displacement value of each vehicle in two adjacent frames of images, if the displacement value of the vehicle exceeding the set ratio is less than the critical value, it is judged as congestion; Extract frame images according to the time series, detect the vehicle target frame of the vehicle position, if the set number of vehicles stop and do not move, it is determined that an accident may occur; if the set number of vehicles stop and are identified as congestion, it is determined that an accident has occurred. 4. The multi-section traffic accident and congestion detection method based on deep learning as claimed in claim 2, characterized in that, the congestion mileage identification is identified by the distance of the congestion camera, specifically, the steps are as follows: Obtain the coordinate information of the camera for the camera that detects congestion; Traverse the cameras adjacent to the congested camera to get all consecutive congested cameras; For cameras that are congested continuously, calculate the distance between two cameras that are congested, and the sum of the distances is the congestion mileage. 5. multi-section traffic accident and congestion detection method based on deep learning as claimed in claim 1, is characterized in that: Congestion accident direction binary classification model, including middle line or barrier identification module, vehicle front identification module and accident direction determination module; Middle line or isolation zone identification module: configured to use the Faster-RCNN algorithm to identify and extract the middle isolation zone or median line; according to the recognition result, determine whether it is congestion on the left or right side of the road; Vehicle front identification module: configured to use the Faster-RCNN algorithm to identify the direction of the vehicle's front according to the images of adjacent frames collected by the camera, and then identify whether it is the direction of the incoming vehicle or the direction of the outgoing vehicle; In chronological order, if the outline of the same vehicle in two adjacent frames of images is getting bigger and bigger, the car in the image is the front of the car, and if the outline is getting smaller and smaller, it is the rear of the car; The accident direction judging module is configured to identify whether the obtained lane is congested on the left or right side according to the current camera orientation, and determines the direction of the congested accident based on the direction of the front of the vehicle. 6. A multi-section traffic accident and congestion detection system based on deep learning, characterized in that it includes: Camera polling control module: configured to use polling mechanism and congestion pre-judgment to sequentially access the camera to obtain vehicle images collected by the camera; Congestion situation identification module: configured to use the trained traffic event three-classification model for vehicle detection and recognition based on the acquired vehicle images to obtain traffic congestion conditions; Congestion accident direction recognition module: configured to obtain vehicle images of traffic congestion road sections, use the congestion accident direction binary classification model to identify, identify the middle line of the lane or the isolation zone, and the direction of the front of the vehicle, and determine the congestion accident direction; The three-category model of traffic events sequentially recognizes the vehicles in the image, the number of vehicles in the image, and the moving distance of the vehicles, and combines the number of vehicles and the moving distance of the vehicles to obtain traffic congestion; The polling mechanism is used to access the cameras sequentially. The dichotomy method is used to predict the possibility of road congestion in each frame of image. The road in the camera's camera area is congested, and continue to acquire the vehicle image data of the current camera; Predict the possibility of smooth roads in each frame of image. The prediction method is to identify the number of vehicles in the picture, set the value of the smoothness confidence coefficient according to the number of vehicles, and judge whether it is smooth according to the value of the smoothness confidence coefficient; Setting the first threshold and the second threshold of the unimpeded confidence coefficient, the first threshold is greater than the second threshold; If the smooth confidence coefficient is not less than the set first threshold, the road in the camera area of the camera is smooth, and the next camera screen is directly switched for analysis; The unimpeded confidence coefficient is between the first threshold and the second threshold, continue to obtain the first set frame number image data of the current camera to calculate the average unimpeded confidence coefficient, determine whether the current road is congested according to the average unimpeded confidence coefficient, and switch Go to the next camera for image acquisition; If the unimpeded confidence coefficient is less than the second threshold, continue to obtain the second set frame number image data of the current camera to calculate the average unimpeded confidence coefficient, determine whether the current road is congested according to the average unimpeded confidence coefficient, and switch to the next camera for image processing collection; Wherein, the first set frame number is greater than the second set frame number. 7. An electronic device, characterized in that it comprises a memory, a processor, and computer instructions stored in the memory and run on the processor, when the computer instructions are run by the processor, any one of claims 1-5 can be accomplished steps described in the method. 8. A computer-readable storage medium, characterized in that it is used to store computer instructions, and when the computer instructions are executed by a processor, the steps described in any one of the methods of claims 1-5 are completed.