Research On Tracking Algorithm Of Surveillance Video Target Based On Twin Network

During the implementation of target tracking algorithm in surveillance video,it is challenging to face complex problems such as illumination change,occlusion between scenes,camera jitter and so on.In 2016,Siam Fc algorithm was proposed,which first applied twin network in the related engineering field of target tracking,and achieved excellent results.In recent years,SiamRPN++ algorithm proposes a new model architecture to perform hierarchical and deep aggregation operations,which improves the accuracy of the algorithm.When SiamRPN++ algorithm uses resnet-50 benchmark network module for convolution kernel operation,convolution operation in each layer of neural network is usually very sensitive to scale changes,which will affect the ability of feature expression.At the same time,SiamRPN++ model has high cost and heavy load in the deployment of monitoring equipment.Aiming at the problems of SiamRPN++,in this paper,under the framework of twin network tracking algorithm,the model is compressed to facilitate deployment on the hardware platform,so as to achieve the effect of improving the speed;then the multi-scale convolution method is used to enhance the CNN core skeleton,so as to improve the robustness in the case of scale changes,and enhance the ability of feature expression.In view of the above ideas,this paper improves SiamRPN++ algorithm in the framework of twin network.The main contents of algorithm improvement are as follows:(1)In the first mock exam,the CNN is used to reduce the computation cost of SiamRPN++ model.The first step is to introduce the SPM(significant SPM)module to each module to achieve the model compression.The pruning method of self-adjusting network model will determine the corresponding pruning strategy according to each layer and each sample,and then compress the whole model.(2)Based on the pruned SiamRPN++ network model,multi-scale convolution operation is used to enhance the core skeleton of Res Net.In convolution operation,a series of expansion rates are combined,and then the expansion rate is redistributed to a single convolution layer in a single convolution core of each convolution layer,and the bottleneck of Res Net-50 is used All the 3×3 standard convolution layers in the middle of blocks are replaced by the improved multi-scale convolution layer,which improves the feature extraction ability of the model without introducing any parameters and increasing the amount of calculation.Finally,it is introduced into SiamRPN++ network model to improve the success rate of feature extraction.In order to verify the effectiveness of the algorithm,the improved algorithm is compared and tested on the OTB100 data set.The results show that the algorithm can achieve performance improvement without loss of robustness and acceptable accuracy.