Research On Visual Tracking Based On Multi-attention Convolutional Neural Network

Visual tracking is an extensively studied problem in computer visions,playing a fundamental role in automatic drive,smart transportation,intelligent manufacturing,military technology and so on.Despite more than 40 years of development and some great achievements,tracking algorithms still need to improve the performance under illumination variation,background clutters,motion blur and deformation.Therefore,further research on visual tracking is of great significance.In order to overcome the shortcomings of the tracker based on multi-domain convolutional neural networks,this paper introduces attention mechanism and optimizes the stages of offline training and online tracking.Finally,a high-precision tracker and a real-time tracker with attention mechanism are constructed.The contributions of this paper can be listed as follows:1.Two types of attention are introduced and embedded in the multi-domain convolutional neural network in an effective manner,allowing the network to adaptively focus on significant features along channel and spatial axes.In addition,attention mechanism increases the diversity of feature extraction and optimizes the feature engineering of the algorithm.2.This paper proposed Attention-wise Global Pooling,and based on this novel method,a light-weighted and effective module named Efficient Dual Attention Module is constructed.It maintains the real-time speed and enhances robustness of tracking algorithm under various conditions through the attention mechanism.3.Inter-instance loss is proposed,which makes full use of the task-driven multi-domain structure to distinguish between instances across different domains.Aiming at alleviating redundant samples during online update,group update strategy for dynamically selecting samples is proposed.Extensive experiments on a standard tracking benchmark demonstrates the effectiveness of our work,and verifies the-state-of-art performance of two trackers in this paper.