| With the gradual commercialization of UAV technology,UAV aerial photography technology has become one of the main means for people to obtain information.It is very important to study the detection and tracking methods of moving objects in UAV aerial video for UAV real-time operation or later video image information processing.However,due to the flexible flight mode of UAV,there are usually complex background,wide viewing angle,variable scale,occlusion loss and other situations in UAV aerial videos,which are all difficulties in the field of image detection and tracking.This thesis mainly studies the detection and tracking methods of moving objects in UAV videos,including video images denoising preprocessing,moving object detection in videos and moving object tracking in videos.The main work and contribution of this thesis are as follows:(1)Aiming at noise may be introduced into the process of acquisition or transmission,which may affect the subsequent target detection and tracking,this thesis,video images denoising preprocessing is used in this thesis.The classic moving target detection algorithms are introduced and compared with experiments,including frame difference method,background difference method and LK optical flow method.The results show that LK optical flow method can fully extract the moving information of the target in the UAV aerial videos with lens jitter,which is convenient for later separation processing.(2)Aiming at the problems of scale change and occlusion loss of moving target in UAV aerial videos,a kernel correlation filter tracking method based on multiscale and target loss retracing is proposed in this thesis,which improves the accuracy of tracking and robustness to target loss.The experimental results show that the improved tracking method has higher accuracy and good real-time performance.(3)In order to make the detection and tracking process more systematic,this thesis carries out the follow-up processing of LK optical flow method,including feature point filtering,rectangular clustering filtering,centroid positioning processing.As a detection module,its output can be used as the initial information of the tracking algorithm.At the same time,the linear kernel correlation filtering algorithm with less computation is used,and the scale adaptive function is added.The method of trajectory prediction is used to predict the possible range after the target is lost,and the detection module is used to re detect the target position within the prediction range,which improves the possibility of retracking the target.The whole part is used as the tracking module.Finally,the systematic process results are analyzed,and good results are achieved. |
