Study Of Vehicle Detection And Tracking In UAV Images Based On Deep Learning

Vehicle detection and tracking are basic tasks of the intelligent transportation system,which can provide import information to alleviate traffic congestion and dispose traffic accident.Most of researches on vehicle detection and tracking are based on videos,which are captured by fixed cameras.Nowadays,with the maturity of the UAV technology,UAVs have been applied in many areas.Compared to fixed cameras,UAVs are flexible,portable,wide field of view and et al.Combing UAVs and intelligent transportation system can extend the region of monitoring and save manpower and material resources.However,compared to traditional fixed cameras,vehicle detection and tracking based on UAV images bring new challenges.Hence,this thesis carried a study of vehicle detection and tracking in UAV images based on deep learning.Aiming at the problem of the existing methods are failure in detecting small-scale vehicles,this thesis proposes a multi-level feature enhanced network for vehicle detection.First,the image is fed into a backbone to generate multiple level feature maps with different resolutions.Then,those feature maps are input to the E-Dnet,respectively.As last,the enhanced feature maps are utilized for classification and regression.In addition,to address the problem of the inequality between confidence score and Io U,which leads to missing vehicles,this thesis designs an Io U-guided loss.The experimental results demonstrate that the proposed method can significantly improve the performance of the detector.In order to deal with the problem that the deep model has many redundant parameters,which leads to require large of computing resource.This thesis proposes a method that combines the model prune and knowledge distillation to accomplish fast vehicle detection.In order to obtain a lightweight network,this method prune a large network.Then,we make the big network as a teacher network and make the lightweight network as a student network.Moreover,this thesis designs a knowledge distillation based on minimum loss strategy,which is applied in object detection.Finally,experimental results represent that the method proposed in this thesis can increase detection speed greatly while maintaining high detection accuracy.For the problem of frequently ID switches in multi-object tracking based on one-shot method,this thesis proposes a method for multi-object tracking based on location constraint.At first,the detector generates detection results including bounding boxes and Re-ID features.The following step is computing the affinity matrix between objects and tracks.We introduce location constraint to computing affinity matrix,which improves final matching results.Experimental results show that the proposed method can observably reduce the number of ID switches and improve tracking stability.