Research On Algorithm Of Visual Tracking Base On Multi-layered Adversarial Siamese Networks

With the development of technology,software and hardware technologies for visual tracking applications have developed rapidly.However,in the complex and varied practical application scenarios,the visual tracking technology also faces many challenges,such as occlusion and deformation of the target.On the other hand,some specific scenarios put forward higher requirements for the processing speed of visual tracking,which also hinders the further promotion of visual tracking.Therefore,designing a tracker for practical application scenarios has an important impetus to the development of visual tracking technology.Deep neural network is a technical tool that has attracted much attention in the field of visual tracking in recent years.This method learns the position of the target in continuous video frames offline or online by learning the tracking process of a large number of tag data,and achieves the goal of target tracking.Throughout the many models of deep neural networks,simple offline networks lack good generalization ability,and the tracking accuracy is often low.For the general online network,real-time adjustment of network parameters causes a lot of time loss,and it is difficult to achieve real-time tracking requirements.This paper focuses on the visual tracking actual application scenario,selects the Fully-Convolutional Siamese Network(Siam-FC)with moderate speed and precision from the mainstream deep network model,and improves from the following two aspects:Aiming at the problem that the input of Siam-FC contains a lot of background,this paper designs a lightweight channel attention algorithm,filters the background features of the target area,and enhances the effective features of different scales with the Inception structure.The model improves the feature extraction ability of the network while reducing the inherent defects in the network.On this basis,the sum-and-max algorithm of different depth features is proposed,so that the final response graph can more accurately locate the target and improve the accuracy of model tracking.The results on the OTB benchmark show that the tracker proposed in this paper has a certain improvement in accuracy.In view of the fact that the visual tracking process may be subject to noise interference,this paper will generate a combination of the Generative Adversarial Networks(GAN)and the Siam-FC,and propose an Adversarial Siamese Network(Siam-GAN).The Siam-GAN includes two parts: a generating network and a discriminating network,wherein the generating network simulates the real environment,and the target is tracked by applying different types of noise interference discrimination networks.Discriminating network learning noise distribution,optimize its own parameters,and overcome the noise impact to accurately track the target.The whole network continuously improves the performance of generating networks and discriminating the network through alternating training.Experiments show that the tracker proposed in this paper has a certain improvement effect on the tracking of targets in complex scenes.