Artificial intelligence-based intersection traffic flow monitoring system

Abstract

Disclosed is an artificial intelligence-based intersection traffic flow monitoring system, comprising: two support columns fixedly disposed at two sides of an intersection, and a cross beam fixedly connected between the two support columns. Multiple cameras are provided on the cross beam, and each camera corresponds to a lane. The cameras perform data exchange with a local processing terminal by means of a wired network or a wireless network. The cameras transmit collected data to the local terminal for processing, the local terminal performs data exchange with a cloud server by means of a remote wireless transmission network, and the cloud server receives image information uploaded by the local terminal, and recognizes and counts vehicles by means of a vehicle recognition algorithm. The cloud server and a control and display terminal perform data exchange by means of a remote wireless network, and information processed by the cloud server is transmitted to the control and display terminal for display. The described system has a high processing speed, and can accurately and quickly recognize corresponding vehicles.

Description

An intersection traffic monitoring system based on artificial intelligence TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle detection, and particularly relates to an intersection traffic monitoring system based on artificial intelligence. BACKGROUND

[0002] With the continuous development of China's economy and the continuous improvement of road traffic construction, real-time monitoring of vehicle flow on highways has gradually become a research hotspot in the field of transportation. On the one hand, urban intelligent transportation has been widely concerned by all sectors of society. Such a stable and reliable system that can monitor the real-time traffic flow of highways 24 hours a day provides a basis for ensuring traffic safety and intelligent control. On the other hand, a convenient real-time traffic monitoring system also provides a reference for the convenience of the general driving and riding personnel. The intersection traffic monitoring system in the prior art has not high enough accuracy in detecting vehicles, which causes the related information processing to be not timely, and thus may cause traffic congestion.

[0003] SUMMARY

[0004] The technical problem to be solved by the present application is to provide an intersection traffic monitoring system based on artificial intelligence which is convenient to use and has high detection accuracy.

[0005] To solve the above technical problems, the technical solution adopted by the present application is as follows: an intersection traffic monitoring system based on artificial intelligence, comprising:

[0006] Two support columns fixedly arranged on both sides of an intersection and a cross beam fixedly connected between the two support columns, a plurality of cameras are arranged on the cross beam, and each camera corresponds to a lane, the cameras perform data interaction with a local processing terminal through a wired network or a wireless network, the cameras transmit the collected data to the local terminal for processing, the local terminal performs data interaction with a cloud server through a remote wireless transmission network, the cloud server receives image information uploaded by the local terminal and identifies and counts vehicles through a vehicle recognition algorithm, the cloud server and a control and display terminal perform data interaction through a remote wireless network, the cloud server transmits the processed information to the control and display terminal for display, and when remote control of the cameras is needed, the control and display terminal uploads a control command to the cloud server through a remote wireless network, the cloud server forwards the control command to a camera driving device through a remote wireless network, and the camera driving device controls the movement of the cameras to complete remote operation of the cameras.

[0007] Further technical solutions are as follows: the vehicle recognition algorithm comprises the following steps:

[0008] Obtaining multiple camera images of different spatial dimensions, and labeling the subcategory and location information of the vehicle on each camera image;

[0009] Performing enhancement processing on each small region picture in each camera image to obtain an enhanced picture, the enhancement processing including random region covering, random adjustment of hue saturation, and random adjustment of brightness and contrast;

[0010] Training a vehicle recognition and counting model using all the enhanced pictures and the labeled subcategory and location information of the vehicle on each camera image, to obtain a trained vehicle recognition and counting model, the vehicle recognition and counting model being a Yolov3 target detection algorithm model based on a MindSpore framework, and the output of the vehicle recognition and counting model after each training being the subcategory and location information of the vehicle;

[0011] When receiving a vehicle recognition and counting request of a city, obtaining a camera image corresponding to the request, and based on the obtained camera image, obtaining the subcategory and location information of the vehicle of the city according to the trained vehicle recognition and counting model, to realize vehicle recognition and counting.

[0012] Further technical solutions are that the method of training a vehicle recognition and counting model using all the enhanced pictures and the labeled subcategory and location information of the vehicle on each camera image, to obtain a trained vehicle recognition and counting model specifically includes:

[0013] Using a Darknet-53 backbone network to extract multi-scale features based on all the enhanced pictures, the multi-scale features being different semantic features of the same entity in different enhanced pictures in different spatial dimensions;

[0014] Using a FPN feature pyramid network to fuse the multi-scale features in different semantics through upsampling and feature fusion, to obtain picture basic features, the picture basic features representing the relative probability between different categories or subcategories of the entity in the enhanced picture;

[0015] Based on the picture basic features, using a yolohead prediction network to predict the subcategory and location information of the vehicle, wherein the location information of the vehicle is represented by a bounding box;

[0016] Using the subcategory and location information of the vehicle predicted by the yolohead prediction network and the labeled subcategory and location information of the vehicle on each camera image to train a vehicle recognition and counting model, to obtain a trained vehicle recognition and counting model;

[0017] When the loss function of the vehicle identification counting model is stabilized at a predetermined threshold, the training is ended, otherwise the parameters of the Darknet-53 backbone network, the FPN feature pyramid network and the yolohead prediction network are updated by back propagation of the loss function, and then the network after the parameter update is trained again.

[0018] The beneficial effects produced by the above technical solutions are that the image collected by the camera is transmitted to the cloud server for processing through the wireless network, the processing speed is fast, the cost is low, and the vehicle at the intersection is identified and counted through the vehicle identification algorithm, the vehicle identification algorithm can accurately and quickly identify the corresponding vehicle, therefore, the system has the advantages of fast processing speed, high identification precision and the like. BRIEF DESCRIPTION OF DRAWINGS

[0019] The application will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

[0020] Fig. 1 is a principle block diagram of the system according to the embodiment of the application;

[0021] Wherein: 1, support column; 2, cross beam; 3, camera; 4, local terminal; 5, cloud server; 6, control and display terminal. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0023] As shown in Fig. 1, the embodiment of the application discloses an intersection flow monitoring system based on artificial intelligence, comprising:

[0024] Two support columns 1 are fixedly installed on both sides of the intersection, and a crossbeam 2 is fixedly connected between the two support columns 1. Several cameras 3 are installed on the crossbeam 2, and each camera 3 corresponds to one lane. The cameras 3 interact with a local processing terminal 4 through a wired or wireless network. The cameras 3 transmit the collected data to the local terminal 4 for processing. The local terminal 4 interacts with a cloud server 5 through a remote wireless transmission network. The cloud server 5 receives the image information uploaded by the local terminal 4 and identifies and counts vehicles using a vehicle recognition algorithm. The cloud server 5 interacts with a control and display terminal 6 through a remote wireless network. The information processed by the cloud server 5 is transmitted to the control and display terminal 6 for display. When remote control of the cameras 3 is required, the control and display terminal 6 uploads the control command to the cloud server 5 through the remote wireless network. The cloud server 5 then forwards the control command to the camera driving device through the remote wireless network. The camera driving device controls the movement of the cameras 3, completing the remote operation of the cameras.

[0025] Furthermore, the vehicle recognition algorithm includes the following steps:

[0026] Acquire multiple camera images from different spatial dimensions, and label the vehicle sub-category and location information in each camera image;

[0027] Enhanced images are obtained by performing enhancement processing on each small area of ​​each camera image. The enhancement processing includes random area coverage, random adjustment of hue saturation, and random adjustment of brightness and contrast.

[0028] The vehicle recognition and counting model is trained using all enhanced images and the vehicle subclass and location information labeled on each camera image. The trained vehicle recognition and counting model is a Yolov3 object detection algorithm model based on the MindSpore framework. After each training, the output of the vehicle recognition and counting model is the vehicle subclass and location information.

[0029] Upon receiving a vehicle identification and counting request from a city, the system acquires a camera image corresponding to the request. Based on the acquired camera image, it uses a trained vehicle identification and counting model to obtain the sub-category and location information of the vehicles in the city, thereby achieving vehicle identification and counting.

[0030] The method for training a vehicle recognition and counting model using all enhanced images and manually labeled vehicle sub-category and location information for each camera image, to obtain a trained vehicle recognition and counting model, specifically includes:

[0031] The multi-scale features are extracted based on all the enhanced pictures by using a Darknet-53 backbone network, and the multi-scale features are different semantic features of the same entity in different spatial dimensions presented in different enhanced pictures.

[0032] The multi-scale features are fused in different semantics by upsampling and feature fusion to obtain picture basic features by using a FPN feature pyramid network, and the picture basic features represent relative probabilities between different categories or subcategories of entities in the enhanced pictures.

[0033] The subcategory and position information of the vehicle are predicted by using a yolohead prediction network based on the picture basic features, and the position information of the vehicle is represented by a bounding box.

[0034] The vehicle recognition and counting model is trained by using the subcategory and position information of the vehicle predicted by the yolohead prediction network and the subcategory and position information of the vehicle labeled on each camera image, and a trained vehicle recognition and counting model is obtained.

[0035] When the loss function of the vehicle recognition and counting model is stable at a predetermined threshold, the training is ended, otherwise the parameters of the Darknet-53 backbone network, the FPN feature pyramid network and the yolohead prediction network are updated by backpropagation of the loss function, and then the network after parameter update is used for training again.

[0036] Preferably, the Darknet-53 backbone network comprises 52 convolutional layers and 1 fully connected layer, the convolutional kernel size in the convolutional layers is alternately composed of 1×1 and 7×7, the step is 3, and group normalization and layer normalization processing are performed between the convolutional layers.

[0037] The yolohead prediction network comprises a first convolutional layer and a second convolutional layer, the convolutional kernel size of the first convolutional layer is 3×3, and the convolutional kernel size of the second convolutional layer is 1×1.

[0038] The images collected by the camera are transmitted to the cloud server for processing through the wireless network, the processing speed is fast, the cost is low, and the vehicle recognition algorithm can accurately and quickly identify the corresponding vehicle, so that the system has the advantages of fast processing speed, high recognition accuracy and the like.

Claims

1. An artificial intelligence-based intersection traffic flow monitoring system, characterized in that... include: Two support columns (1) are fixedly installed on both sides of the intersection and a crossbeam (2) is fixedly connected between the two support columns (1). Several cameras (3) are installed on the crossbeam (2), and each camera (3) corresponds to a lane. The camera (3) interacts with the local processing terminal (4) through a wired network or wireless network. The camera (3) transmits the collected data to the local terminal (4) for processing. The local terminal (4) interacts with the cloud server (5) through a remote wireless transmission network. The cloud server (5) receives the image information uploaded by the local terminal (4) and identifies and counts vehicles through a vehicle recognition algorithm. The cloud server (5) and the control and display terminal (6) interact with each other via a remote wireless network. The information processed by the cloud server (5) is transmitted to the control and display terminal (6) for display. When it is necessary to remotely control the camera (3), the control and display terminal (6) uploads the control command to the cloud server (5) via the remote wireless network. The cloud server (5) then forwards the control command to the camera driver via the remote wireless network. The camera driver controls the movement of the camera (3) to complete the remote operation of the camera (3).

2. The intersection traffic flow monitoring system based on artificial intelligence as described in claim 1, characterized in that, The vehicle recognition algorithm includes the following steps: Acquire multiple camera images from different spatial dimensions, and label the vehicle sub-category and location information in each camera image; Enhanced images are obtained by performing enhancement processing on each small area of ​​each camera image. The enhancement processing includes random area coverage, random adjustment of hue saturation, and random adjustment of brightness and contrast. The vehicle recognition and counting model is trained using all enhanced images and the vehicle subclass and location information labeled on each camera image. The trained vehicle recognition and counting model is a Yolov3 object detection algorithm model based on the MindSpore framework. After each training, the output of the vehicle recognition and counting model is the vehicle subclass and location information. Upon receiving a vehicle identification and counting request from the city, acquire the camera footage corresponding to the request. The system uses head images and, based on the acquired camera images, obtains the sub-categories and location information of vehicles in the city according to a trained vehicle recognition and counting model, thereby achieving vehicle recognition and counting.

3. The intersection traffic flow monitoring system based on artificial intelligence as described in claim 2, characterized in that, The method for training a vehicle recognition and counting model using all enhanced images and manually labeled vehicle sub-category and location information for each camera image, to obtain a trained vehicle recognition and counting model, specifically includes: Using the Darknet-53 backbone network, multi-scale features are extracted from all enhanced images. These multi-scale features are the different semantic features of the same entity in different spatial dimensions in different enhanced images. By utilizing the Feature Pyramid Network (FPN), multi-scale features are fused with different semantics through upsampling and feature fusion methods to obtain basic image features. These basic image features represent the relative probabilities of entities in the enhanced image involving different categories or subcategories. Based on the basic features of the image, the YOLOhead prediction network is used to predict the vehicle subclass and location information, where the vehicle location information is represented by bounding boxes; The vehicle recognition and counting model is trained using the vehicle subclass and location information predicted by the YOLOhead prediction network and the vehicle subclass and location information labeled on each camera image. Training ends when the loss function of the vehicle recognition and counting model stabilizes at a predetermined threshold; otherwise, the parameters of the Darknet-53 backbone network, FPN feature pyramid network, and YOLOhead prediction network are updated through backpropagation of the loss function, and then the network with updated parameters is trained again.

Images