Target Tracking Method Based On Deep Learning And Motion Feature

Target tracking plays an important role in the computer vision industry,which automatically locate a target in subsequent frames according to the state of a presupposed target in the initial frame.Research results on target tracking have been widely used in the fields of medical imaging,military and intelligence surveillance,autopilot systems,and human–computer interaction.Most basic studies on target tracking have focused on the several basic aspects: feature extraction,target apparent modeling and tracking strategies.The effectiveness of target tracking largely depends on the feature extraction,and the proper target features will be the key to successful tracking.Currently,the deep learning is commonly used for feature extraction,and a common structure is the Siamese network model.However,existing tracking methods based on deep learning feature extraction tend to train offline and extract the apparent information of the target in a single frame for tracking,while ignoring the fact that the tracking video also possesses sequence features.The target position relationship before and after the video sequence can be used to exclude the influence of distractors in complex environments.To this end,SiamCAR network framework is employed in the proposed model,and the marginal distribution,recursive OTSU method are used to determine whether strong interference appears in the search region of the current frame.If strong interference appears,the motion features of the target are used to exclude the interferers in the search window and improve the performance of the tracking algorithm.The main content of this article includes the following two aspects:(1)Considering the problems of similarity interference,partial occlusions,and scale variations during target tracking,a target tracking method based on SiamCAR combined with kalman filtering is proposed.Under the proposed framework,the marginal distribution of the feature maps is used to determine the presence or absence of interferents.When interference is present in a scene,a motion vector composed of the predicted value obtained through a Kalman filter is used as the basis for target prediction.Experiments show that SiamCAR with motion features achieves the best performance in videos with similar object interference,partial occlusion,fast motion,and small target tracking.(2)A tracking method based on SiamCAR combined with LSTM is proposed.Under the proposed framework,the recursive OTSU algorithm is used to derive the extreme value points of the response feature map that determines whether strong interference occurs.When interference is present in a scene,a motion vector composed of the predicted target center obtained through a LSTM module is used as the basis for target prediction.Experiments on the benchmark LaSOT dataset show that the proposed algorithmachieves the best performance in videos with similar object interference,partial occlusions,and small target tracking,compared with the classical SiamCAR and other excellent target tracking algorithms.