Research On Video Object Tracking And Recognition Algorithm Based On Deep Siamese Network Model

With the continuous advancement of computer science and technology,target tracking technology has also developed rapidly.Nowadays,target tracking is an important research direction in the field of computer vision,which is one of the current research hotspots.Its technology has been widely used in video automation monitoring,unmanned driving,motion analysis,and video detection.Therefore,the research on the video target tracking algorithm has important practical significance.It is extremely challenging to complete the tracking and recognition of video targets in a complex environment,while attaches importance to both real-time and accuracy.Through in-depth analysis of the tracking results of the classic Siamese tracker(Siam FC)in different scenes,it is found that it is difficult to handle scenes with background clutters and motion blur.In response to this problem,this paper uses a deeper backbone network VGG-16 to make the target representation more efficient and accurate,and combines locality-constrained linear coding with similarity measures to improve the accuracy of target template and search image matching.This paper starts from the convolutional layer of the Siamese network backbone architecture,decomposes the standard convolution into multiple convolutions and builds an asymmetric convolution block,in order to strengthen the feature representation ability of the network and reduce the network's dependence on the structure.To improve the representation ability of the network for occlusion targets,this paper proposes a Feature Drop Block(FD)based on the framework of Siamese network,which can not only simulate the occlusion in the real world,but also alleviate overfitting to a certain extent,so as to can effectively improve the tracking effect of occluded targets.Finally,the proposed asymmetric convolution block and FD are embedded into two representative Siamese network trackers for experimental verification.The algorithm in this paper has been compared with other excellent tracking algorithms on the public data sets OTB100 and VOT2018,which suggested that the algorithm in this paper can achieve higher accuracy and speed on the deep Siamese network model that compared with other algorithms.