Research On Video Target Tracking Technology Based On Deep Learning

Reviving the development of computer vision,the research on video target tracking technology has always been one of the discussed hottest topics in IT industry.By extracting and analyzing the context information of the images in the video sequence,the target tracking can predict the motion state of the target and calibrate the position of the target.Target tracking not only extends the rich data resources in the field of target tracking scenario analysis,but also provides powerful help for correct detection and recognition of moving target.Nevertheless,the video target tracking technology is still not mature.In the process of tracking,the target will not only be disturbed by occlusion,blur and changes in posture,but also may be affected by the natural environment,such as changes in lighting and weather.All of these interference factors may lead to poor tracking effect.Therefore,the research of video target tracking technology is still a challenging work.Aiming at the two problems existing in the Siamese-FC tracking algorithm: the accuracy and speed of tracking can not reach the relative balance state and the target is easily affected by the factors such as occlusion or complicated background,which may leads to the problem of target being lost or tracking drift.In this paper,two new target tracking algorithms are proposed that based on the theory of deep learning and Siamese neural network.Which mainly does the following two aspects: First,a Siamese-SE network tracking algorithm is proposed.Specifically,the SE-Network substructure is embedded in the network model,and the correlation between each channel in the image is analyzed and modeled.The purpose is improving the characterization ability of features and reducing the calculation amount to save the calculation time,so that the speed and accuracy of tracking can reach a balanced state relatively.Second,a Siamese network tracking algorithm based on temporal attention mechanism is proposed.Specifically,adding the temporal attention mechanism to the network model,firstly,collecting the positive samples and negative samples that from the current frame and positive samples that from the historical frame,and the convolution operation is performed on the samples.Then the temporal attention mechanism module is used to give adaptive weights to the convoluted features.Finally,the state of the predicted target is updated online so as to reduce the influence of occlusion or complicated background on the result of tracking.In this paper,three different public benchmarks were trained and tested,and compared with a variety of advanced tracking algorithms based on traditional machine learning and based on deep learning.The experimental results were analyzed by using the evaluation criteria such as Success Rates,Average Overlap and OPE(One Pass Evaluation).The results show that the algorithm that we proposed is superior to other comparative algorithms in tracking performance,and has good robustness for a variety of tracking interference factors.