基于改进YOLOV3的商场人员类型及流量分析

Shopping Mall Personnel Type and Traffic Analysis Based on Improved YOLOV3

摘要: 随着人口数量在全世界范围内快速增长，对特定场景如商场中人口数量的控制就成为了重要的研究课题。近年来，公共场所当中因为人流密集和大量拥挤而产生的对社会稳定有所影响的事故屡见不鲜。传统的人力控制方法费时费力，近年来随着人工智能和高清摄影机技术的不断完善，利用计算机对商场实行监控成为了主流趋势。然而目前的探头技术仍然只能实现观察和监测的作用，无法对一段监控视频中的人流进行数量上的计算。本文意在实现对商场人员的人流进行数据化以及可视化的监测。与此同时，尽可能地对商场的特定时段的不同人员类型进行分析，这样不仅能很好地维持商场秩序和人流量控制，同时还为商场管理人员提供了思路，可以通过对不同时段的经营策略进行针对化的调整，来更好地对商场进行管理和经营。YOLO算法是一种经典的目标检测算法，本文将使用较新的YOLOV3算法对商场的人员类型和流量进行检测，建模以及分析。并给出了多个维度的模型参数的性能指标图，以此来表明该算法的可行性。

Abstract: With the rapid growth of population around the world, the control of population in specific scenarios such as shopping malls has become an important research topic. In recent years, accidents affecting social stability in public places due to the dense flow of people and mass crowding are not uncommon. The traditional manual control method is time-consuming and laborious. In recent years, with the continuous improvement of artificial intelligence and high-definition camera technology, the use of computers to monitor shopping malls has become the mainstream trend. However, the current probe technology can only realize the role of observation and monitoring, and can not calculate the number of people in a surveillance video. The purpose of this paper is to realize the digital and visual monitoring of the flow of shopping mall personnel, at the same time, as far as possible to analyze the different types of personnel in the specific period of the shopping mall, so as not only to maintain the order of the shopping mall and the control of the flow of people, but also to provide ideas for the shopping mall management personnel, through the adjustment of the business strategy of different periods of time, to better manage and operate the shopping mall. YOLO algorithm is a classic target detection algorithm. In this paper, the relatively new YOLOV3 algorithm will be used to detect, model and analyze the personnel types and traffic of shopping malls. The performance index graphs of model parameters with multiple dimensions are given to show the feasibility of the proposed algorithm.