Research Of Mean Shift Target Tracking Algorithm Based On Spatio-temporal Visual Saliency Features

Target tracking has always been a hot spot in computer vision research and has been widely used in many fields such as national defense,intelligent transportation and living security.In recent years,the study of visual saliency detection technology has made substantial progress and breakthrough,the technology can make the image processing process closer to the human cognitive mechanism.Some scholars have successfully introduced it into the target tracking task to improve the robustness and accuracy of the tracking algorithm.However,the existing visual saliency models are not effective enough in detecting complex scenes and do not have universality.In order to further improve the performance of the tracking algorithm,this paper improves the saliency detection model based on the contrast of the histogram,and studies a saliency detection model which is suitable for the target tracking task.The model is introduced into Mean Shift based target tracking algorithm,and a Mean Shift target tracking algorithm based on spatial-temporal visual saliency features is proposed.Specific research includes the following two aspects.Firstly,aiming at the problem that the saliency detection model based on histogram contrast is not pertinent when dealing with tracking tasks,and ignores the contribution of motion information to saliency,a spatio-temporal saliency detection model suitable for target tracking task is proposed.In the spatial domain,the spatial saliency detection model based on histogram contrast is improved by introducing SLIC superpixel segmentation and prior information of target position,and the complete and saliency target of each frame is obtained.In the time domain,the motion feature channel is added as the guidance method of the visual attention,the motion vector is detected by the frame difference method and the optical flow method,the saliency of the motion vector is analyzed by using the features of motion entropy and direction consistency,and then the time-domain salient model is constructed.Finally,saliency maps of time domain and spatial domain are merged using adaptive weighting approach to obtain the spatio-temporal saliency detection model that can adapt to different situations of tracking tasks.Experimental results on standard datasets show that the proposed model has better robustness than contrast models,and can effectively restrain disturbing background disturbances and highlight the target.Secondly,aiming at the problem that the traditional Mean Shift based target tracking method is not robust enough when the colors of target and background are similar or the target is occluded,this method is improved by using the spatio-temporal saliency detection model proposed in this paper,the adaptive adjustment strategy of convergence window and Kalman filtering position prediction are introduced at the same time,and then a Mean Shift target tracking algorithm based on spatial-temporal visual saliency features is proposed.The proposed algorithm can give full play to the invariance of visual saliency features in the disturbed environment such as light changes and posture changes,and then improve the tracking robustness and accuracy.The experimental results show that the visual saliency features can effectively compensate for the lack of color features and improve the robustness of tracking,and the algorithm can achieve continuous adaptive tracking under the conditions of color confusion and partial occlusion,comparing with contrast algorithms,this algorithm has stronger anti-blocking,anti-interference ability.