Research And Application Of Target Tracking Based On Siamese Networks

With the emergence of video data information like seawater,the popularization of artificial intelligence,and the target tracking technology has been widely used in military and civilian applications.Researchers at home and abroad are paying more and more attention to the field of target tracking.When using target tracking technology for tracking,it is necessary to determine the initial state of the target and other prior information.However,due to the lack of prior information of the target,and the unpredictable changes of the target itself and the external scene in the subsequent video sequence,will lead to the degradation of tracking performance.Therefore,it is significant and challenging to study a tracking algorithm with high accuracy,small computation and strong robustness.In the past,due to the limited level of video image processing technology,correlation filtering algorithm used manual features in target tracking.At present,with the development of target tracking technology,the features obtained by deep learning technology have strong expression ability.Researchers combine the target tracking algorithm with deep learning technology,and the tracking performance is generally improved.The tracking algorithm based on siamese networks,with its excellent performance,has gradually come into the attention of researchers.Among them,Siam FC is the most typical target tracking algorithm.After that,researchers have extended the Siamese networks tracking algorithm.DenseSiam uses densenet as the benchmark networks to improve the performance of the algorithm,and the overall performance is outstanding in many deep networks models.When applying the DenseSiam network model for target tracking in intelligent video surveillance scenarios,the following problems will arise: As the network density increases,the amount of model calculation and internal redundancy will increase sharply,and the deployment cost of hardware devices such as intelligent video surveillance is higher;The model cannot make full use of the similarity relationship between the input samples when distinguishing positive and negative samples.Therefore,in view of the above problems,two optimization methods are proposed:(1)The model compression of the benchmark networks densenet is carried out by combining weight pruning and quantization grouping convolution,which reduces the redundancy of features and the consumption of calculation,and compresses the model size and calculation amount within the acceptable range.(2)In order to solve the problem of insufficient use of correlation among input samples,the loss of triples is added to the model.The existing triple loss function is improved to enhance the discriminative ability of positive and negative samples and the discriminative ability of the model,which makes the tracking performance more robust.In order to verify the performance of the algorithm,the improved algorithm is tested on OTB2015 data set.The experimental results show that the improved algorithm has improvement in tracking accuracy and success rate under stable tracking speed.