The Research Of Object Tracking Method Based On Convolution Neural Network

With the development of artificial intelligence,object tracking technology has been rapidly developed in the fields of human-computer interaction,intelligent monitoring,automatic driving and so on.Although the object tracking algorithms based on deep learning have achieved good results in the field of tracking,target occlusion,fast moving,scale change and target deformation are still the main problems to be solved urgently.In this paper,based on the most popular Siamese networks in the current object tracking field,an object tracking algorithm based on residual attention mechanism and fusion spatiotemporal context information is proposed.In order to overcome the problem of poor robustness and precision of the GOTURN(Generic Object Tracking Using Regression Networks)tracking algorithm,this paper proposes an object tracking algorithm that combines time and space information.The algorithm transmits the target template,prediction region and search region to the network to extract the general feature maps and predict the tracking target position in the current frame through the fully connected layers.At the same time,the residual attention mechanism network structure is added to the target template network structure to enhance the feature representation ability of the network and improves the overall performance of our algorithm.Aiming at the problem of inadequate feature extraction from images based on Siamfc(Fully-Convolutional Siamese Networks)network,a target tracking algorithm based on multi-scale dilated convolution is proposed.Based on the Siamfc,two multi-scale dilated convolution network blocks are added to the current fully-convolutional Siamese network tracking algorithm.By using different convolution kernels and dilated convolution layers with different dilated rates,the feature maps of different scales and different receptive fields are obtained,and then these feature maps of different scales are fused together to enhance the network's feature expression ability.This method effectively improves the overall performance of the original tracking algorithm in all aspects.A large number of experiments on the current mainstream tracking test data sets show that the performance of the original tracking algorithm is significantly improved by the proposed two tracking algorithms.It not only improves the target occlusion and multi-scale problem,but also improves the tracking performance in other complex scenarios.