Object Tracking Based On Multi-domain Convolution Neural Network Combined With Prior Knowledge

Object tracking,as one of the hot research fields of computer vision,has been concerned by domestic and foreign researchers,and the related research in the top conferences and periodicals is common,and the research results play a central role in the field of missile guidance,autonomous driving and other automation.In recent years,algorithms based on correlation filter and deep network have become the mainstream algorithm of object tracking.Multi-domain convolution neural network(MDNet)tracking algorithm,as a representative of deep network object tracking algorithm,won the championship in VOT2015 object tracking competition,has high success rate and accuracy,but because of the low resolution of target training sample,lack of samples,easy to lose in the case of background similarity interference,and the deep feature repeat extraction and the network frequent training update,consumes more computing resources,the speed is slow,and it can't be used in practice.To solve the above problems,this paper proposes an algorithm framework combining correlation filter location priori and multi-domain convolution neural network to improve the robustness of the algorithm,and at the same time optimize the network structure and update strategy to improve the speed of the algorithm.The main research work of this paper is as follows:(1)A multi-domain convolution neural network object tracking algorithm combining position priori is proposed.The algorithm uses the prediction result of high performance correlation filter algorithm(spatial regular correlation filter,SRDCF)as a priori information and input into deep network,the deep network detect the reliability of the prediction target position of correlation filter,and uses the prediction position as the benchmark to remove the interference candidate samples from the deep network according to the sample deviation level.The remaining candidate samples of the highest confidence level are selected as the target,and finally the filter and deep network model are updated according to the confidence of the tracking results.The algorithm is evaluated and tested in OTB-100 and VOT-2016 datasets,and the experimental results show that the tracking performance of this algorithm is higher than that of MDNet algorithm,and the effect is improved significantly under the condition of background similarity interference.(2)A new network structure(FastNet)and network update frequency Adaptive strategy are proposed.Firstly,the ROI-Align layer is added after the convolution feature extraction layer of the original network to solve the problem of repeated extraction of convolution features of the previous target candidate samples,which greatly improves the efficiency of feature calculation,and at the same time,in order to alleviate the problem of imprecise target positioning caused by the feature pooling of ROI-Align layer,the pool layer after the second convolution layer is removed.The third convolution layer adopts dilated convolution,the resolution of the feature graph is increased one times,the semantic information contained is enhanced,the target state is measured according to the confidence level of each frame target,the frequency of network update is adjusted adaptively,and the speed is further improved.(3)Improve the DSST correlation filter model and use its prediction results as prior information.The network confidence level is introduced into the DSST model,and the target state is judged by PSR: when the two are above the threshold,the filter is updated every two frames,and when the PSR is below the threshold and the confidence level is higher than the threshold,the filter is forced to update with the deep network prediction result,and when both are below the threshold,the filter stops updating.The whole algorithm is measured under OTB-2013 dataset,the performance is better than the mainstream real-time tracking algorithm,the speed is 25 FPS,and the real-time video tracking requirement is achieved.