Research On Object Localization And Tracking System For UAV

With the advantage of flexibility and efficiency,Unmanned Aerial Vehicles have many applications,such as automatic navigation,traffic monitoring and rescue.Computer vision is one of the most important technology for these implementations.In this thesis,the task of object localization and tracking in computer vision is studied and applied to extract position and motion state information of ground targets in aerial images transmitted by the airborne image transmission system.The main contents and research results of the thesis are summarized as follows:(1)The vehicle detector for a drone is built based on the improved single-stage detection framework of RefineDet.The receptive field block is used to improve the backbone network of RefineDet,which uses different convolution kernels and dilation ratio to simulate the receptive field of human visual cortex.Results tested on the UAV vehicle dataset show that the mean Average Precision(mAP)of the proposed model is 5.8% and 1.9% higher than that of Yolov3 and RefineDet,respectively.(2)The SiamRPN tracking framework is applied to track static markers in aerial videos.The distractor-aware module and the attention mechanism are used to improve the robustness and accuracy of the SiamRPN model.Video sequences with static targets in the UAV123 dataset are used for evaluation.Experimental results show that the overlap score and Area Under Curve(AUC)of the proposed model are 19.29 and 6.94 higher than those of the SiamRPN model.(3)The tracking problem of moving targets is more complicated than the static target tracking problem,facing more challenges such as deformation,occlusion and out of view.The proposed network consists of offline candidate regression network and online verification network.The candidate regression network is built based on the SiamRPN framework to generate candidates,which is improved with the CIR module to maintain more localization information and extract effective features.The verification network based on MDNet is used to verify candidates,and the PSROIAlign layer and the dilated convolution are applied to improve the speed and accuracy of the verification network.The tracking system combines the output of these networks to determine whether to re-detect the target.Experimental results videos with moving targets in the UAV123 dataset show that the Precision,overlap score and AUC of the proposed model are 17.52,9.68 and 15.24 higher than those of SiamCIR-LT.(4)We use DJI Phantom 3 to collect aerial videos with building(static target)and pedestrian(moving target)to test the actual performance of the proposed trackers.Experimental results show that both trackers can keep track of the target with high accuracy.