面向注意力增强和特征选择的多目标跟踪算法  

Multi-Object Tracking Algorithm Based on Attention Enhancementand Feature Selection

作　　者：殷向阳 李树辉 陈禹铭 文峰[1] YIN Xiangyang;LI Shuhui;CHEN Yuming;WEN Feng(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159

出　　处：《沈阳理工大学学报》2022年第4期26-31,共6页Journal of Shenyang Ligong University

基　　金：辽宁省教育厅高等学校基本科研项目(面上青年人才项目)(LJKZ0267)。

摘　　要：一阶段的多目标跟踪算法具有速度快的优势,但存在目标检测质量低、身份标识号(ID)切换次数过于频繁的问题。为此,提出一种基于注意力增强和特征选择的多目标跟踪算法。通过增加平行重识别(Re-ID)分支完成特征提取任务;通过设计空间注意力和通道注意力机制的方式降低特征图的噪声,提升特征图的质量;通过加入特征选择模块,提取检测特征图和Re-ID特征图。经测试集验证,该方法在提升精确度、降低ID切换次数方面均取得了进展,提出的注意力增强和特征选择的方法可以明显提升目标跟踪效果。One-shot multi-object tracking algorithms have been widely studied for their advan-tages of high speed.However,such methods have some drawbacks of low quality of object de-tection and frequent ID switching.To address these problems,a novel multi-object tracking al-gorithm based on attention enhancement and feature selection is proposed.The algorithm ex-tracts feature by adding parallel Re-ID branches.The noise of feature map can be suppressed and the quality can be significantly improved by the design of spatial attention and channel at-tention mechanism.At the same time,detection feature map and Re-ID feature map are extrac-ted by adding the feature selection module.The results of test datasets show that the proposed method completely has made progress in improving accuracy and decreasing the number of ID switching.Thus,the proposed method of attention enhancement and feature selection can sig-nificantly improve the tracking effect and achieve efficient tracking of objects.

关 键 词：目标跟踪 目标检测 注意力增强 特征选择 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]