Research On High-speed Object Detection And Tracking Based On Deep Learning

A large number of high-speed fragments are formed after the high explosive in the warhead exploded.These fragments destroy targets by penetrating,jetting,etc.The flight states and flight parameters of fragments are obtained by fragments detection and tracking,which are essential basis for studing combat power of warheads.General image processing methods can hardly satisfy image processing and measurement requirements in the high-speed photography scene,for fragments with small size,large number and high speed are difficult to be detected in complex and diverse shooting range environments.Therefore,the research of fragment swarm detection and tracking based on deep learning has important engineering application value.The main contents and contributions of this thesis are as follows.Fragments with small size and large number are difficult to be detected,which results in detection errors such as false detection and missed detection.To solve the problems,a fragment detection algorithm based on multi-scale fusion YOLOv4 network is proposed in this thesis.Based on the YOLOv4 network,a shallow feature map sensitive to small targets is added to the original three feature layer scales,and multi-scale feature fusion is used to detect fragments at four scales.Kmeans ++ clustering algorithm is also utilized to obtain priori boxes suitable for fragments,which improves the accuracy of fragment detection.Compared with the original algorithm,the experimental results validate the performance of the proposed scheme with an increase of more than 8% in the detection accuracy,recall rate and average recognition rate.Occlusion of moving fragments may lead to target loss and tracking errors between two consecutive frames.Consequently,a fragment tracking algorithm is proposed based on multi-feature fusion Deep Sort algorithm.The the candidate matching method is used in the proposed algorithm.Through multi-feature fusion,CNN features and HOG features are utilized to match fragments in turn.If any feature matches successfully,the Kalman filter and the Hungarian algorithm are used to obtain the trajectory of the matched target.Otherwise,cascade matching is utilized for the last match,if the match still fails,the trajectory of the fragment is considered as a new trajectory.The experimental results prove that the MOTA and MOTP of the proposed scheme are improved by 6.4% and 2.8% respectively.Based on the results of fragment detection and tracking,the speed and scatter angle of the fragments are calculated,and the evolution process of fragment swarm is deduced to provide indispensable parameter basis for studying variation of the attacking ability of warhead fragments.In order to realize the visualization of fragment detection and tracking,a fragment target detection and tracking system is built by PYQT,which provides an important technical means for warhead design and damage effectiveness evaluation.