Research On Forward Multi-vehicle Target Tracking Method Based On Deep Learning

Environmental awareness is a crucial step for driverless vehicles and the basis for other modules of driverless vehicles.Its ability directly determines the working conditions of driverless vehicles.Visual multi-target tracking based on deep learning is a key technology in current environmental perception.It can track the surrounding targets of vehicles stably,solve the problem of vibration caused by discontinuous target detection,and provide data support for vehicle trajectory prediction.The thesis is based on the target tracking section in the project "Cooperative Sensing and Target Tracking Technology for Complex Road Environments",which mainly focuses on the front multi-vehicle target tracking method based on deep learning,and studies the two mainstream multi-target tracking methods.Compared the advantages and disadvantages of commonly used databases for intelligent vehicle research,analyzed the application scenarios and usage methods of the database,and have chose the database required for each training.Based on the analysis of various evaluation indicators of multi-target tracking,the vehicle multitarget tracking evaluation index is selected on the basis of combining the research method and characteristics of vehicle multi-target tracking.For the detection-based multi-target tracking method,introduced the benchmark DeepSort model,based on the original model,to optimize the learning of the deep cosine metric,and use the center loss function and the cross entropy loss function together to replace a single cross entropy loss function to increase classification accuracy;Retrain the CNN-based re-identification module using the VeRi vehicle dataset.The detection module uses Gaussian yolov3,and uses the processed KITTI data set for binary classification training to improve the target detection accuracy.The detector output is used as the input of the tracking model.Compared with the original,the tracking accuracy and accuracy are significantly improved.For the multi-target tracking method that integrates single targets for tracking,the benchmark model used is the DeepMOT model.Based on a full understanding of the Seq-to-Seq BiRNN network,SiamRPN ++,which performs better in the twin network,is used as the single-target tracker.GRUs with fewer parameters and faster training speeds serve as the cyclic unit of Bi-RNN.The training data set uses the UA-detrac processed for labeling,and the test data set uses the output of the EB detector.Comparing the output of the changed model with the original model,the tracking accuracy is significantly improved.