Research And Application Of Multi-Vehicle Tracking Algorithm With Distance Information Based On End-to-End

With the rapid growth of the number of motor vehicles in the country,the number of on-street parking lots will inevitably increase,and the on-street parking lot needs to use a high-level camera parking system to track vehicles and handle events such as vehicles entering and exiting parking spaces.At the same time,the system can also use the distance information between vehicles to better determine the position of the vehicle and the state of the parking space.However,there are many problems in the vehicle multi-object tracking algorithm in this scene at this stage,such as ID switching caused by occlusion.Based on the above background and existing challenges,the thesis conducts study and research.The focus of the thesis is to improve the existing two-stage multi-object tracking algorithm.Obtaining distance information between vehicles while tracking is another research point of the thesis.The thesis considers to improve the one-shot multi-object tracking algorithm model,and the distance measurement adopts the monocular bird’s-eye view distance measurement.The main work is as follows:(1)In order to improve the speed of the whole system,first change the detectionbased tracking to end-to-end tracking(that is,the detection branch is parallel to the ReID branch),select the FairMOT algorithm framework,replace the detection head with YOLOX,and add cross-stage Partial Network(CSP)and Path Aggregation Network(PANet).In order to ensure the accuracy of the system,according to the different key information of the detection task and the Re-ID task,the attention mechanism is introduced,the spatial attention module is added to the detection branch,and the channel attention module is added to the Re-ID branch.Finally,the data association algorithm is improved,combining feature information and motion information,and taking different operations for targets with different confidence intervals.The thesis self-made the evaluation data set in the field scene to verify the above improvements.The speed of processing each frame increased by 5ms,and both MOTA and IDF1 increased by about2%.(2)The theories related to ranging are studied,the bird’s-eye view ranging algorithm based on the monocular camera is selected,and the coordinate system and the conversion between the coordinate systems related to the bird’s-eye view ranging algorithm are studied.Afterwards,the camera is self-calibrated to obtain the camera’s internal reference matrix and external reference matrix.After obtaining the parameters,select four sets of corresponding point coordinates,and use the theoretical knowledge of linear equations to calculate the transformation matrix from the imaging plane to the bird’s-eye view plane,complete the conversion from the imaging plane to the bird’s-eye view,and then complete the distance measurement.In the thesis,the algorithm is verified and evaluated in the experimental environment.After verification,the error is within 5cm,which meets the actual requirements.After that,the algorithm is applied to the scene of high-position camera parking on site,which is convenient for later judgment on vehicle status and parking space information.(3)In the thesis,the vehicle multi-object tracking algorithm is combined with the ranging algorithm,and the algorithm is deployed on the JetsonTX2 embedded platform,which improves the correctness of the system by 17%,and uses the TensorRT library for acceleration to ensure that the processing speed reaches at least 10 per second frame.Later,a Pyqt-based visual assistance system was developed in the background to solve possible problems in the system with the help of vehicle events and distance information(for example,the ID switching caused by the long-term occlusion of the vehicle can be corrected by distance information),and finally show the whole system.result.