Research On Moving Object Detection And Tracking Based On Robot Vision

In recent years,moving object detection and tracking technology has been widely used in the field of robot vision.The research in this direction is not only the premise of dealing with high-level visual tasks,but also involves many disciplines.Therefore,it has always been the focus and difficulty in the field of vision research.In this dissertation,the video collected by camera is used to simulate the vision of robot,and the algorithm of moving object detection and tracking is mainly studied.In order to improve the accuracy of detection and tracking,the work conducted is as follows:1)In terms of the problems related to "holes" and shadow false detection of moving foreground objects obtained by Gaussian mixture model detection algorithm,an improved detection algorithm is proposed.Firstly,the three-frame differencing method combined with Gaussian mixture model is used to extract the motion region.Secondly,Canny algorithm is introduced to detect the edge of the results obtained by three-frame differencing method to obtain the object contour.Finally,the "And" operation is performed on the moving area and the object contour,and then the final moving object is obtained through morphological processing.The experimental results show that the improved algorithm can effectively solve the problems of "holes" and shadow false detection and improve the detection accuracy.2)In terms of the situation that the object tracking algorithm is easy to fail under the interference factors such as occlusion and scale change,an improved object tracking algorithm based on DSST and particle filter is proposed.A particle filter tracking module and an adaptive discrimination module are added to the DSST tracking algorithm.The adaptive discrimination module is used to judge whether the object tracking drifts occur in real time,and the particle filter based on color distribution is introduced to predict the position of the object after tracking drifts.In order to ensure the accuracy of tracking,the scale filter in DSST algorithm is retained to determine the object scale,and the tracking model is updated according to the tracking results.The experimental results show that the accuracy and success rate of the improved algorithm are 86.6% and 81% respectively,and it shows good tracking performance in various interference scenes,especially after the object is occluded.3)Aiming at the need of multi-object tracking in practical application of robot vision,a multi-object detection and tracking algorithm based on YOLOv3 and kernel correlation filtering is proposed.The trained network model is used to detect pedestrian objects,the detected objects are assigned a fixed ID,each object is input into the tracking module in parallel,and KCF algorithm is used to track each object separately.In addition,a correction strategy is proposed to update the object position and number of tracking modules,thus improving the accuracy of the algorithm.The experimental results show that the proposed algorithm can realize multi-object tracking with a success rate of 81%,which meets the requirements of practical application.