Visual Tracking Algorithm Based On Deep Feature Sharing With Key Point Matching And Tracking

As an important research direction in computer vision,visual target tracking technology has been deeply and widely applied in many fields,such as autonomous driving,urban traffic,human-computer interaction,AI(Artificial intelligence)medical service,etc.However,due to complex background environment of the actual scene as well as the uncertainty of the object and several other uncertain factors,most traditional tracking algorithms cannot achieve desired results.Although algorithms based on deep learning can meet the requirements of high tracking accuracy,it is difficult to complete target tracking with a high execution time.Therefore,the key of target tracking is how to deal with the appearance change of objects and occlusion adaptively.Based on the above analysis,we presented a feature extractin method based the depth image features extracted by convolutional neural network and the traditional low-level image features in this paper.Then,high-level depth convolution features which have both image semantic information and spatial structure information with traditional low-level feature which have high resolution are combined in order to propose a visual tracking algorithm based on deep feature sharing with key point matching and tracking,which enables the object to be located quickly and accurately in addition to having a good ability of target representation.The following are the main research contents and innovations of this paper are described in this following:(1)Tracking frame innovationWe present a tracking framework based on traditional artificial design features and highlevel depth features to handle the task of object tracking.Firstly,the depth features extracted from the offline network are shared.Then,the target detection with appropriate feature point matching algorithm is utilized.After that,the detection module the tracking module are combined together to estimate the object position.This new tracking framework not only obtains the comprehensive description of high-level semantic features and low-level features,but also achieves the double guarantee of online tracking and traditional tracking in the selection of tracking model,which establish a solid foundation for long-term stable tracking.(2)Introducing an adaptive template update mechanismA template updating strategy based on the variation trend of feature points in the target region is designed,which can realize adaptive template updating according to the appearance information of target on the basis of avoiding the introduction of background and other error information.This mechanism makes the algorithm more adaptable to the appearance change of targets,and avoids the high complexity,high computation and object drift in the situation of updating tracking parameters frame-by-frame.(3)Propose a target occlusion processing mechanismBy dividing the object area into blocks reasonably and combining with the information change of each sub-block in the continuous frame image,a new anti-occlusion processing mechanism based on block correlation is proposed,to quickly and accurately adapt to target occlusion.Finally,the problem of occlusion can be handled in long-term tracking.