Study On Moving Object Detection And Tracking Algorithm Based On UAV Vision

The Moving object detection and tracking technology is the hotspot of current research in computer vision,which has been widely used in medical,transportation and military fields.With the development of UAV technology,the quad-rotor UAV has strong maneuverability,great concealment and high flexibility.There is a new breakthrough when the visual tracking combines with the UAV,this technology has been extensively used in traffic control,crop protection and unmanned aerial vehicles,and so on,so it has important research value and practical significance.The information of moving object which is obtained by the UAV vision system can be dealt with tracking algorithm in order to accomplishing object tracking by UAV.In this paper,the correlation filter-based tracking algorithm is applied to quad-rotor.Firstly,introduced the UAV flight principles and hardware structure design,the visual tracking system of UAV,unmanned machine lens correction and video processing process is given.Secondly,the saliency detection algorithm is used in the pre-processing of dealing with object information.Introduced how to extracting the features of the saliency detection,and the evaluation criterion of the saliency detection algorithm.This paper explains the principle of several classical saliency detection algorithms in detail,and gives the experimental comparison.The improved MGR algorithm is proposed,which is more robust,showed in the experiment.Thirdly,as there are many of tracking algorithms which are based on correlation filter,general tracking steps and framework of this algorithm are summarized in this paper,and the standard of target tracking evaluation is also given.Explained the MOSSE,CSK,STC and KCF algorithm principle in detail,and given the experimental results,experiments show that the KCF algorithm is the most superior,not only fast but also good tracking effect.Fourthly,in order to eliminate the redundant information of the background in KCF algorithm,utilized the saliency detection algorithm to process the initialization: initialized the target in the first frame as input image to the saliency detection in order to separate the target and background,according to the actual size of the target re calibration of the tracking frame size and position.Experimental results show that using this algorithm could get a more compact tracking box after preprocessed,and eliminate the most of irrelevant background information.Fifthly,in order to increase the multi scale function of the KCF algorithm,proposed an improved Mean Shift algorithm combination with KCF algorithm in this paper: used the filter response peak to predict the target position,and the improved Mean Shift algorithm to recalculate the size of tracking box.Experiments show that the modified KCF has good performance on multi scale.