Research On Deep Learning Based Single Target Video Tracking Algorithm

Visual target tracking technology is a research hotspot in the cross field of computer vision and pattern recognition,which is widely used in military and daily life.Although researchers at home and abroad have made a lot of achievements in the field of target tracking in recent years,because video is affected by blur,occlusion,rotation,light change and other factors,it is difficult for the tracker to track the target stably for a long time,so target tracking is still very challenging.In the traditional target tracking algorithm,the algorithm based on the correlation filter framework greatly improves the speed of the tracking algorithm by introducing the Fast Fourier Transform,but the accuracy of the tracking algorithm needs to be improved.With the emergence of massive video data and the improvement of computing power of computing equipment,deep learning technology has been greatly developed and widely used in target tracking algorithm.Among them,the introduction of convolution neural network and residual network model further improves the performance of target tracking algorithm.In this paper,the single target tracking algorithm based on convolutional neural network is improved by optimizing the network structure of feature extraction and training loss function of improved model.The main contributions are as follows:(1)A target tracking algorithm based on residual attention mechanism and region overlap rate is proposed.On the basis of the traditional feature extraction network,a special network based on the residual attention mechanism is added to generate the target region saliency map,weaken the background and highlight the foreground,and guide the tracker to pay attention to the parts that are helpful to the results.In addition,in the network structure of feature extraction,the deep and shallow features of the network are fused in a cascade way to further improve the discrimination of features.In the later stage of network model training,the loss function based on area overlap rate is introduced,which makes the algorithm model obtain better positioning effect.The experimental results on vot2016 dataset show that the accuracy of the improved algorithm is 2.76% higher than that of the baseline algorithm.(2)A twin network tracking algorithm based on multi-level feature fusion is constructed.On the one hand,based on the original feature extraction network structure,multi-level feature discrimination structure is introduced to synthesize the output results of different levels of features,so that the output accuracy of the network is higher.On the other hand,on the basis of the original network model training,the balance term loss is introduced to balance the influence of a large number of simple negative samples in the training samples,so as to improve the network model tracking effect.The experimental results on otb2015 data set show that the improved algorithm achieves 78.4% average distance accuracy and 73.6% average coincidence accuracy.