Object Tracking System Based On Siamese Network

Nowadays,with the development of the network,all aspects of people are constantly integrated into the network.The advent of 5G era also makes people urgently need to enjoy the gains brought by artificial intelligence computer vision.Target tracking is an important research direction in the field of computer vision and pattern recognition.It is widely used in the fields of video automatic tracking and monitoring,automobile automatic driving and military weapons.But in real life,the scenes are often very complex and diverse.The video is filled with obstacle occlusion,serious deformation of the target,moving too fast,background interference,too narrow observation field and changes in illumination environment,which make the original tracking algorithm unable to track the target in real time,accurately and robustly.With the development of computer technology and the improvement of hardware performance,convolutional neural network has been well applied in tracking field.The proposal of full convolution twin in 2016 adds new directions to target tracking and gives birth to a large number of new algorithms.This paper mainly studies the tracking problem of single target moving object.Using the idea of twin network,starting from the full convolution twin network based on convolution neural network,some improvements have been made to the twin network in view of its low accuracy,inadequate real-time,background interference drift and other problems.The main contents of this paper include the following two aspects:(1)An adaptive tracking algorithm based on twin networks is proposed.One of the twins is responsible for the recognition and judgment of the appearance information of the target,and the other is responsible for the feature extraction of the semantic information around the target.Semantic branch is Pre-trained AlexNet network,while appearance branch is based on full convolution twin network by adding some pixel-level information features,and adding a feedback interruption network between the first three layers of network,so that the processing of simple frames can be tracked quickly when encountering simple frames.Because appearance information belongs to shallow information,when appearance branch network interrupts in advance and the whole network gets feedback,the network will automatically choose to shield semantie information to track the target directly,so as to improve the real-time performance of the network.Otherwise,the target location can be obtained by superimposing the two network features in a certain proportion.(2)A model of attention channel mechanism based on VGG twin network is proposed,which uses the first ten layers of VGG-16 network as the basis and adds attention channel to the target line.The appeal information of VGG twin network can better capture the characteristics of the target.Different functional channels have different effects on tracking in different scenarios.Channel weights represent the different importance of different channels.Therefore,the adaptability of the algorithm can be improved by increasing the weights of more stable characteristic channels.