Research On Target Tracking Algorithms Based On Joint Correlated Filtering And Model Generation

Target tracking has become one of the basic research topics in the field of computer vision because of its wide application in video surveillance,human-computer interaction,and driverless vehicles.Although great progress has been made over the past decade,target tracking is still a tricky subject due to factors such as illumination variation,deformation,local occlusion,fast motion,and background clutter.In order to solve these problems in the tracking process,our work mainly includes the following two aspects:(1)The Mean Shift(MS)algorithm has a good effect on partial occlusion,but the target model only uses color features,lacks target space information,and has insufficient description of the target,which is prone to error location.Kemelized Correlation Filters(KCF)have a poor tracking effect on occlusion and fast motion,but their circular matrix describes the target more fully.In the tracking process,when the size of the target changes,tracking the target with a fixed template will lead to an increase in background clutter when the target is too small,while the extraction of target features will be reduced when the target is too large.Both of the above algorithms use fixed templates to track targets,in view of their advantages and disadvantages,a scale adaptive algorithm based on mean shift and kernel correlation filtering is proposed to solve the problem of target tracking under fast motion,motion blurring and scale change.(2)When the target is partially occluded or rotated,the Discriminative Scale Space Tracker(DSST)algorithm will completely lose the target position and drift.Mean shift algorithm will jitter when the target is partially occluded,but it can track the target again after ocelusion.Mean shift algorithm uses eolor histogram to deseribe the target model.The color histogram can be used to record the probability of the color appearance of the target,so that it is not affected by the rotation of the target,and finally track the object accurately.In addition,if color features are fused into DSST algorithm,the robustness of the algorithm can be further improved.Therefore,combining the advantages and disadvantages of both,a scale adaptive target tracking algorithm based on mean shift and multi-feature flusion is proposed.Experiments show that the algorithm achieves good tracking results on datasets with partial occlusion and object rotation.