Research On Target Tracking Method Based On Deep Learning

The field of computer vision is mainly composed of target detection,target recognition and segmentation,target tracking and other research directions.Among them,the research of target tracking is the focus in the field.With the development of deep learning,hardware systems,big data and cloud computing,the research of target tracking technology has also gained new opportunities.However,due to the fact that target tracking only gives the first frame of target boundary box and lacks the data information provided to the deep learning network,it is not widely used in computer vision as other fields.Therefore,this paper starts from two aspects: target detection and target tracking,and explores the target detection and tracking based on deep learning.In this paper,a target detection method based on the Faster Regions-Convolutional Neural Network(Faster R-CNN)with pruning channels and fully convolution is proposed.Firstly,the method uses LASSO regression to prune redundant channels on each channel of the CNN in order to accelerate feature extraction.Then,the linear least square method is used to reconstruct the minimum error to reduce the impact of pruning channels on the network.Finally,the fully convolutional VGG-16 network is used to share the computation of ROI in order to accelerate the reasoning time.In this paper,a spatio-temporal regularized correlation filter for object tracking method based on two-branch Siamese fully convolutional network learning is proposed.Firstly,a two-branch Siamese network framework is constructed,where the semantic branch network is deep semantic features and the appearance branch network is the appearance features.Then,the spatio-temporal regularized correlation filter layer is added into the semantic branch to improve the tracking speed and the tracking accuracy.Finally,the semantic branch network and the appearance branch network are combined to further optimize the tracking performance.Experimental results conducted on Caltech,VOT2017 and OTB2015 dataset demonstrate that the proposed target tracking method has better robustness and tracking performance under complex background,occlusion and illumination change.