基于非局部感知网络的运动目标跟踪方法

Nonlocal Neural Network-Based Moving Target Tracking Metho

针对跟踪运动目标过程中网络对目标被遮挡或目标周围存在干扰物敏感,从而导致不可靠的响应位置和错误跟踪框的问题,提出一种基于深度学习的免锚框孪生卷积网络跟踪方法。首先,通过非局部感知网络来学习目标引导的特征权重,该权重用于细化目标模板分支和搜索分支的深度特征,以监督的方式利用两个分支特征的远程依赖性,从而有效抑制噪声干扰。其次,进一步开发一个包围框感知块将多维回归特征与跟踪质量相关联,这个模块加强目标模板分支和搜索分支之间的相互作用,提高网络定位准确性。在标准数据集上的实验结果表明,所提方法能实时跟踪目标,并在准确度上获得提升。

Typically,networks are sensitive to targets being blocked or interference around the target when tracking moving targets,resulting in unreliable response positions and incorrect tracking frame.Thus,an anchorfree Siamese networktracking approach based on deep learning is proposed.First,the feature weight of the target guidance is derived through the nonlocal perceptual network,which is then applied to refine the depth features of the target template branch and search branch,and to improve the remote dependence of the two branch features in a supervised manner to effectively suppress noise interference.Second,to correlate the multidimensional regression features with the tracking quality,a bounding box perception block is developed.This module strengthens the interaction between the target template branch and the search branch and enhances the accuracy of network positioning.Furthermore,the proposed method can track the target in real time and enhance accuracy,according to the experimental findings on standard data sets.