Research On Object Tracking Method Based On Deep Multi-task Learning

In recent years,visual object tracking has been widely studied in the field of computer vision.Visual object tracking technology is widely used in video surveillance,and has been applied in smart city,traffic control,public security monitoring,safety inspection and other aspects.Filters combined with deep feature show excellent performance in visual tracking,which greatly affects the development of object tracking.Many existing tracking methods only use the depth feature of the target to separate the features of the deep semantic information of the object,and fails to explore the relevance of the depth feature,which limits the strong representation ability of the deep feature to a great extent,resulting in the low robustness of the tracking algorithm.In response to the above problem,this paper focus on the following works:(1)A multi-task deep dual correlation filters for visual tracking(MDDCF)method is proposed.A new multi-task learning(MTL)tracking method is proposed to make the best of the multi-level features extracted from deep networks,which takes the target representation with individual features as a single task to better explore the correlation between different deep features.The alternating optimization method of the objective function is re derived and proposed to train the correlation filter and network parameters.An effective model updating scheme is used to capture the appearance changes of the target.Experiments on benchmark datasets show that the proposed method has good performance in the attributes of object occlusion and fast motion.(2)A multi-task convolution operators with object detection for visual tracking(MCOT)method is proposed.In order to optimize the convolution filter,multi task learning technology is used to explore the correlation between continuous deep features.An updating strategy combined with weighted filters is introduced,which integrates weighted filters into tracking model to reduce the update times of parameters for speeding up object tracking.Object detection network is introduced to solve the problem of target loss in the framework of object tracking,to improve the robustness of the tracking method.Experiments on tracking data sets show that the proposed method has good performance in the attributes of long-time target loss and fast motion.