Research On UAV Video Object Detection And Tracking Algorithm

Unmanned Aerial Vehicles(UAVs)are suitable for video image collection in areas where information collection is difficult due to their small size and strong mobility.Nowadays,the application range of drones is becoming more and more extensive.From the beginning of the military field to the various civil fields afterwards,there are shadows of drones assisting people in their work.Target detection and tracking tasks,so the research on the detection and tracking of targets in the UAV scene plays an important role in many areas of people’s lives.However,due to the special location and viewing angle of the UAV itself for image collection,it brings many difficulties to target detection and tracking.For example,when the UAV’s flying altitude is high,it will cause the tracking target in the video to be too small;When the angle of view of the airborne lens changes,it will cause irregular deformation of the target;when the environment where the drone is tracking the target is complex,the target will often be blocked.In this paper,aiming at these outstanding problems,the target detection and tracking algorithm in the UAV scene is researched.The following is the specific research content of this paper:.1.Carry out specific analysis and experimental research on target detection in the UAV scene.The algorithm ideas and network architecture models of three object detection algorithms based on deep neural networks are introduced,namely the FasterRcnn algorithm based on the one-stage target detection idea and the YOLOv3 and Retina Net algorithms based on the two-stage target detection idea,and apply them in the image scene collected by the drone,a comparative experiment was carried out on a drone data set,and the existence of these three neural network models in the detection of targets in the drone scene was analyzed in detail.Problems and performance.2.A correlation filtering target tracking algorithm based on re-detection is proposed.Since there are often problems such as target occlusion in the scene of drones,and once the target is occluded,the original tracking algorithem will drift.The method constructs a re-detection mechanism.When the offset or the target reappears after being occluded,locate the correct target position in time,and conduct a lot of experiments on multiple data sets to compare the performance of this method and other common target tracking methods.Experiments show that proposed tracking method has both improved success rate and accuracy,especially in terms of success rate.In addition,it also shows that the ideas of propsed method can be applied to many other types of target tracking algorithms.3.Aiming at the practical application of UAV target detection and video target tracking algorithms,combined with the research results of this subject,based on the redetection-based video target tracking algorithm and deep learning detection network,the design and realization of video target detection and Track the prototype system.