基于多目标跟踪的商场热点图生成方法
来源:用户上传
作者:郑舟轩 王勇 王瑛
摘要:多目标跟踪领域以基于检测的跟踪方法为主,CenterTrack算法提出了以目标中心点为检测对象,每帧输出基于目标中心点生成的热图以辅助下一帧的检测和跟踪的方法。此方法在保证帧率的前提下有效提升了多目标跟踪准确率,但由于其缺乏对目标重识别的关注,当目标遭遇到遮趸蛟肷影响从检测结果中丢失时无法将随后重新出现的同一目标识别为原目标,导致ID切换较频繁。该文在CenterTrack算法模型中加入近期丢失跟踪链队列和重识别模块以改善其在重识别方面的表现。输入商场监控录像并取得跟踪结果后,根据行人目标移动与停驻时间分配权值生成商场热点图,帮助经营者提升销售能力。
关键词: 计算机视觉; 多目标跟踪; CenterTrack算法; 目标重识别; 商场热点图; MOTA(多目标跟踪准确率)
中图分类号:TP391 文献标识码:A
文章编号:1009-3044(2021)36-0106-03
开放科学(资源服务)标识码(OSID):
Shopping Mall Heat Map Generation Method Based on Multi-object Tracking
ZHENG Zhou-xuan, WANG Yong, WANG Ying
(School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China)
Abstract:The field of multi-object tracking is based on tracking-by-detection methods. The CenterTrack algorithm proposes a method that uses the center of the object as the detection target, and outputs a heat map generated based on the center of the target for each frame to assist the detection and tracking of the next frame. This method effectively improves multi-object tracking accuracy under the premise of ensuring the frame rate. However, due to its lack of attention to target re-identification, when the target encounters occlusion or the influence of noise, thus lost from the detection result, it cannot reappear as the same object. The object is not recognized as the original object, which leads to frequent ID switching. This paper adds a recently lost tracks queue and re-identification module to the CenterTrack to improve its performance in re-identification. After the input of the shopping mall surveillance video data and obtaining the tracking results, the shopping mall heat map is generated according to the weights of pedestrian target movement and staying time to help shop owners improve sales.
Key words:computer vision; multi-object tracking; CenterTrack algorithm; target re-identification; shopping mall heat map; MOTA(Multi-Object Tracking accuracy)
有店铺零售业指的是消费者购买行为主要在一个相对固定,能够进行商品陈列、展示和销售的场所或设施中进行的零售业务[1]。随着网络零售业为首的无店铺零售业的迅速发展,有店铺零售业受到了前所未有的冲击,尤其过去两年受新型冠状病毒肺炎疫情等事件影响,在中国零售业整体增速持续减缓的背景下,有店铺零售业的发展更是连续受挫。尽管有店铺零售业的“退潮”似乎在所难免,许多零售业经营者并未轻易放弃,而是积极寻求改变。
在许多高出货量或者高客单价的零售店铺经营工作中,经营者常常通过观看店铺的监控录像,记录并统计各个货架、陈列台前停驻的消费者数量,以了解店铺内各货架、陈列台的热门程度,并根据这些信息调整货架摆放、商品陈列策略,改善经营状况最终得到更高的经济效益。
多目标跟踪(Multi-Object Tracking, MOT)指的是使用计算机视觉技术,通过处理输入的视频获得视频流中的多个目标以及它们的外观特征、每帧的位置等信息,并最终得到所有目标的运动轨迹[2]。随着深度学习的发展和深入应用,多目标跟踪领域的研究进步迅速,Zhou等提出的CenterTrack在MOT17数据集上进行的实验达到了在22FPS的帧率下,67.8%的多目标跟踪准确率(Multi-Object Tracking Accuracy),截至原文发表时为该公开数据集最佳成绩。多目标跟踪的迅速发展让一些零售业经营者看到了一个能够帮助他们改善经营状况的契机:用计算机视觉技术代替上述手动进行的热门货架和陈列台的记录、统计工作,通过对店铺的监控录像进行处理,以消费者为目标进行多目标跟踪任务,自动统计消费者在各个货架、陈列台前的逗留时间,并据此自动生成相应的商场热点图。
转载注明来源:https://www.xzbu.com/8/view-15425223.htm