Automatic Vehicle Detection And Tracking In Aerial Images By Fixed-wing UAV

Target detection and tracking is an important research field in the computer vision area.The automatic detection and tracking of unmanned aerial vehicle(UAV)ground targets plays an important role and strategic need in military and early warning fields.In these applications,the ground target is usually small in the field of view.The existing detection algorithm is still relatively weak for small targets,especially in the initial stage of tracking.Because of the distance,targets seems too small(target size is less than 10 × 10 pixels)to detect,and only the edge information and a little or even missing texture information exist.In addition,the ground target detection and tracking for fixed-wing UAV also face problems such as complex background environment,fast field of view,interference,etc.This paper mainly studies the automatic detection and tracking system of ground vehicle targets for fixed-wing UAV.This paper first introduces the research and development status of target detection and tracking system for UAV,and try to find out the shortcomings of the current technologies.At the same time,this paper analyzes the development trend of target detection and tracking algorithm.A new target detection method based on target properties is proposed.Candidate targets are extracted by the saliency detection combined with gradient response.After adding the local sensitive histogram(LSH)feature,it can adapt well to the variable lighting conditions outdoor.The size of candidate target in the corresponding pixel position of the field of view is coupled with the attitude information of the fixed-wing UAV,and fake targets can be effectively eliminated.The joint rotation invariant feature,FHOG + LBP,is introduced to make full use of the valuable features of the small target and a very small number of texture features.The problem of the target detection is transformed into the detection of the candidate target,and the pixel-by-pixel search is greatly reduced to achieve the real-time algorithm.With difficult sample mining during training stage,the accuracy increases 10.23%.This paper tries to apply convolution neural network(CNN)in small target detection,and analyzes the shortcomings and its reasons of current deep learning methods.Negative sample mining and re-training are also added during training stage.Compared with the improved Faster R-CNN,accuracy increases nearly 40%.In additioin,an deep learning method based on target properties has been proposed.It achieves 85.34% AP on VEDAI dataset.After target detection,KCF is used for the real-time tracking of the small target.The fixed-wing UAV adopts the embedded ARM platform,and the proposed algorithms are transplanted based on the embedded linux to realize the real-time automatic detection and tracking of the ground vehicle target.The vehicle detection test is carried out by the fixed-wing UAV in complex background environments,and state-of-the-art algorithms are compared on VEDAI,which is a public data set.In order to verify the robustness of our system,the proposed algorithms are simulated through test flying of our fixed-wing UAV in a variety of scenarios.At the same time,the proposed algorithms are compared with the current state-of-the-art algorithms in a public data set.The proposed small targets detection algorithms,including proposals based on target properties(PBTP),joint rotation invariant feature FHOG + LBP and the neural network methods for small target detection,and spatio-temporal context tracking method based on KCF,have been simulated and experimented.Finally,all the related algorithms have been transplanted into ARM platform.