Research And Application Of Single-target Joint Tracking And Segmentation Algorithm

Visual object tracking is one of the basic research directions in the field of computer vision,which aims to track and estimate the scale of any given object in the video.Object tracking technology has been widely used in the fields of auto-driving,robot,intelligent monitoring and military defense.In recent years,the development of deep learning technology has greatly improved the performance of object tracking algorithm.However,face to the challenges of similar interference,non-rigid deformation and drastic scale changes,there are still problems of poor tracking robustness and low accuracy,which are difficult to meet the performance requirements of actual applications.In this paper,two object tracking method are proposed based on deep learning.Firstly,in order to alleviate the problem of model drift and low accuracy,this paper proposes a tracking algorithm based on center-distance between regionals weighted overlap predictor.Secondly,in order to further improve the robustness and accuracy of tracker,a dual correlation filter location fusion mechanism tracker is proposed in this paper.The detailed information is as follows:The tracker based on center-distance between regionals weighted overlap predictor divides the tracking task into three sub tasks: object location,scale estimation and object segmentation.In the object location,correlation filters are used to locate the object;in the scale estimation task,distance and IOU are both used to predict the quality score of dynamic anchors.In order to track the target and obtain the mask of object fine segmentation at the same time,the object segmentation network is added into the tracking framework,and the robustness and accuracy tracking and segmentation of the object is realized by end-to-end multi-task learning.The proposed tracker was evaluated on two tracking benchmark VOT2018 and OTB100.Experiments shows that the proposed tracker achieves EAO index 0.459,2 percentage points higher than DIMP(ICCV2019),robustness index 0.169,accuracy index 0.617 on the VOT2018,and 0.677 success rate index and 0.877 accuracy index on OTB100.Compared with the previous state-of-the-art trackers,the proposed tracker achieves better performance.Furthermore,the proposed tracker is applied on UAV infrared video object tracking,and the effect is remarkable.In order to further reduce the redundant and improve the robustness and accuracy of object tracking,a dual-correlation filter location fusion mechanism tracker is proposed in this paper.The tracker simplifies the object tracking task into two sub tasks: object location and object segmentation,which removes scale estimation.In object location,two independent correlation filters are used to locate the object and fusion location map,which can overcome the influence of similar object interference,and has higher fault tolerance than single filter location,which can effectively improve the robustness of tracker;in object segmentation,a feature modulated encoder-decoder segmentation network is proposed,which can predict fine mask of object directly by combining with object location information.The mask improves the tracking accuracy,delete the rectangular box regression branch and reduces the calculation redundancy.The proposed tracker is fully tested on VOT2018,VOT2019 and GOT10 k.The EAO index of VOT2018 is 0.467,VOT2019 is 0.334,and the average overlap index on GOT10 k dataset is60.0.Compared with the advance tracker,the proposed tracker has more competitiveness.