基于对目标理解和感知的检测跟踪算法  

Detection and tracking algorithm based on target understanding and perception

作　　者：袁武飞 李伟光 熊兴中 黄铃轩 YUAN Wufei;LI Weiguang;XIONG Xingzhong;HUANG Lingxuan(Artificial Intelligence Key Laboratory of Sichuan Province(Sichuan University of Science&Engineering),Yibin Sichuan 644000,China)

机构地区：[1]人工智能四川重点实验室(四川轻化工大学),四川宜宾644000

出　　处：《计算机应用》2022年第S02期67-71,共5页journal of Computer Applications

基　　金：四川轻化工大学人工智能重点实验室项目(ZX123)。

摘　　要：传统的核相关滤波器(KCF)目标跟踪算法利用目标的纹理特征进行相关运算定位,没有检测目标类别,因此在目标纹理被噪声干扰,例如目标运动模糊、快速抖动、目标遮挡等情况时的跟踪精度和成功率较低。针对这些问题,提出一种语义分割和多特征融合相结合的目标检测跟踪算法。该算法将目标跟踪分为检测和跟踪两个部分:在检测阶段使用全卷积网络(FCN)语义分割对场景进行语义分析,对场景中的目标进行分类;在定位阶段使用KCF算法进行跟踪定位,为了提高跟踪精度,将目标的方向梯度直方图(HOG)特征和颜色(CN)特征融合为新的特征。在标准数据集OTB-100视频序列上的实验结果表明,相较于KCF算法,所提算法的跟踪精度和成功率分别提高了14.3个百分点和13.2个百分点,有效提高了跟踪性能。The conventional Kernel Correlation Filter(KCF)target tracking algorithm uses the texture features of the target for correlation operation positioning without detecting the category of the target,resulting in low tracking accuracy and success rate when target texture is disturbed by noise,such as the target motion blur,fast shaking and target occlusion.To address these problems,a target detection and tracking algorithm combining semantic segmentation and feature fusion was proposed.The algorithm divides target tracking into two parts:detection and tracking.In the detection section,semantic analysis of the scene was performed by using fully Convolutional Network(FCN)semantic segmentation,and the targets in the scene were classified;in the tracking section,the KCF algorithm was used for tracking and positioning.In order to improve the tracking accuracy,the Histogram of Orientation Gradient(HOG)features and Color Name(CN)features of the targets were fused into new features.Experimental results on the standard dataset OTB-100 video sequences show that compared with the KCF algorithm,the tracking accuracy and success rate of the proposed algorithm are increased by 14.3 percentage points and 13.2 percentage points,respectively,which effectively improves the tracking performance.

关 键 词：核相关滤波器 目标跟踪 目标检测 语义分割 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]