Research On Single Object Tracking Based On Deep Learning

Object tracking has received increasing attention over the past decades.From particle filter framework to correlation filter,the performance of object tracking algorithm is gradually improved.In this paper,the single object tracking based on deep learning are studied.New solutions are proposed for some problems,and experiments are carried out on related datasets to verify:(1)An object tracking method based on feature maps fusion Siamese network is proposed,to solve the shortcomings of the full convolution Siamese network for object tracker.The method optimizes the process of feature extraction of Siamese tracker.It extracts features of each layer separately to avoid losing spatial information after multi-layer feature extraction.From the deep layers of Siamese network to the shallow layer,the feature maps are fused step by step,and then a similarity learning is conducted to get the final response map.Multi-level features are fully leveraged through a feature transfer block,further improving the discriminability of proposed method using both high-level semantic and low-level spatial information.Experimental results show that the proposed method has high accuracy.(2)A comparative method object tracking based on response maps fusion Siamese network is proposed.The full convolution Siamese network for object tracker formulate tracking as convolutional feature cross-correlation between a target template and a search region.This tracker runs in real-time.However,when there is similar interference,Siamese tracker still has an accuracy gap.Therefore,an object tracking method based on response maps fusion Siamese network is proposed.Different from the full convolution Siamese network for object tracker,this method propose a layer-wise feature aggravation structure for the cross-correlation operation,which helps the tracker to predict the similarity map from features learned at multiple levels.Moreover,This method further propose a depth-wise separable correlation structure which not only greatly reduces the parameter number in the target template branch,but also stabilizes the training procedure of the whole model.The fusion response maps can effectively avoid the loss of spatial information after multi-layer feature extraction.In extensive experiments,proposed tracker consistently show tracking performance and runs in real-time.In this paper,two improved methods for deep learning based single object tracking are proposed,and has carried out experimental verification on related datasets,and finally obtained good tracking accuracy.