基于机器视觉的地铁站客流自动检测方法  

A Method of Automatic Pedestrian Counting in Metro Station Based on Machine Vision

作　　者：陈艳艳[1] 陈宁[1] 周雨阳[1] 赖见辉[1] 张伟伟[1] 

机构地区：[1]北京工业大学建筑工程学院,北京100124

出　　处：《公路交通科技》2013年第10期122-128,共7页Journal of Highway and Transportation Research and Development

基　　金：交通运输部重大科技专项(2012-364-220-109

摘　　要：基于机器视觉的地铁站客流自动检测技术目前在检测精度和稳定性上距实际应用要求仍存在一定差距。针对这一问题,提出了一种适应于地铁站的客流自动检测新方法。该方法针对目前地铁站监控设备的普遍布设位置,研究了基于近景、非垂直视野的客流自动检测算法。该方法首先通过采集行人头部的Haar特征,利用由AdaBoost算法训练形成的强分类器进行人头检测,然后使用双线性插值算法,针对图像中的不同位置,以宁缺毋滥的原则,去除尺寸过大或过小的误检框。最后根据人头检测目标,设计行人的追踪统计算法。该算法由目标聚类、断帧合并和行人驶离3部分组成。试验结果显示,采用该方法进行地铁站内的客流自动检测可进行有效的方向区分,统计精度稳定在90%左右,基本满足实际的应用需求。Now it is difficult to use the method of automatic counting pedestrians in metro station based on machine vision because of its low precision and poor stability. To solve this problem, we proposed a novel method of automatic counting pedestrians common in metro station. Our method achieved the function of pedestrian recognition suitable for short-shot and non-direct downward view for the general layout of cameras in metro stations. First, this method collects the Haar features from pedestrian head samples and detects the pedestrian head using strong classifier trained by AdaBoost algorithm. Second, considering different locations in image, this method removed the mistake detection which is too big or too small using bilinear interpolation algorithm according to the principle of put quality before quantity. Finally, according to the detections of pedestrian head, we designed the pedestrian tracking algorithm which includes pedestrian object clustering, break frame-sequence merging and pedestrian leaving. The experimental result shows that using our method in metro station to count people automatically is very effective. Its accuracy reached 90%, so it can be applied in practice.

关 键 词：交通工程 客流检测 机器视觉 行人 

分 类 号：U491.116[交通运输工程—交通运输规划与管理]