Research On The Detection And Tracking Algorithm Of Multiple Target Vehicles In Dense Scenes

As an important research foundation for autonomous driving,multi-target detection and tracking of vehicles has been widely used in different fields.In dense scenarios with a large number of vehicles and pedestrians,the occlusion and overlap between vehicles seriously affect the accuracy of multi-target detection results and the precision and stability of multi-target tracking,which has become a research hotspot and difficulty.The specific work and innovation points of this paper are as follows:(1)Research on multi-object detection algorithm in dense scenarios.Aiming at the reduction of detection accuracy of multi-target vehicles in dense scenes,an improved algorithm DA-YOLOv5 is proposed.The C3 layers in the backbone feature extraction network is modified in this algorithm.The CBAM lightweight attention module is introduced in the C-C3 layers,which promotes the feature extraction ability and reduces the loss of effective information.Then the classification and regression branch lines are separated by decoupling the network detection head to increase the regression efficiency and accuracy of the algorithm effectively.Finally,the Alpha Io U loss function is used to replace the original CIo U loss function for the further improvement of the regression accuracy.Considering the imbalanced of three types of vehicle samples in the COCO vehicle dataset,the traditional method of combining data enhancement and Grid Mask occlusion method is adopted to enhance the dataset and sample complexity with few samples.In addition,the experiments are verified on the enhanced dataset.The experimental results show that compared with the original YOLOv5 algorithm,DA-YOLOv5 algorithm converges faster,and the accuracy is improved by 2.14%.It can be seen in the visual comparison experiment that DA-YOLOv5 algorithm is more robust in vehicle detection in different scenarios,and can effectively avoid the phenomenon of false detection and omission.(2)Research on multi-object tracking algorithm in dense scenarios.In order to address the problems of low tracking accuracy and unstable tracking model for multi-target vehicles in dense scenarios,an improved DeepSORT multi-target tracking algorithm based on DA-YOLOv5 method is proposed.In this improved algorithm,the Res Ne St50 network is used as the appearance feature extraction network,which realizes the feature extraction across the feature graph groups and enriches the feature information of the channel dimension.Instead of the original constant noise matrix of the Kalman filter,the noise covariance matrix based on the detection confidence is introduced to the improved algorithm.It realizes the noise scale adaptation to obtain more accurate motion state and improve the accuracy of the Kalman filter update.Finally,DA-YOLOv5 model is used as the detector of the tracking model to upgrade the confidence of the detection results.The experimental validation result on the UR-DETRAC dataset show that compared with the original DeepSORT algorithm,the accuracy and precision of the proposed multi-object tracking algorithm respectively increased by 1.04% and 1.66%,and the number of identity switching decreased by 52.99%.In the visual comparison experiment,the improved multi-object tracking algorithm can continuously and correctly track the vehicle target,which effectively boost the accuracy and stability of the algorithm.