地铁应用场景中半监督群体异常行为检测模型
程媛1吴帆2张宁2徐炜2
Detection Model of Semi\|supervised Group Abnormal Behavior in Metro Application Scenario
CHENG YuanWU FanZHANG NingXU Wei
-
作者信息:1.1.中铁第四勘察设计院集团有限公司，430063，武汉；
2.2.东南大学智能运输系统研究中心轨道交通研究所，210018，南京
-
Affiliation:China Railway Siyuan Survey and Design Group Co., Ltd., 430063, Wuhan, China
-
关键词:
-
Key words:
-
DOI:10.16037/j.1007-869x.2023.05.004
-
中图分类号/CLCN:U29\|39
-
栏目/Col:学术专论
摘要:
为了解决拥挤、奔跑及踩踏等地铁群体异常行为带来的公共安全风险问题，针对地铁车站监控视频样本的种类繁多、均衡性差、背景环境复杂多样等问题，提出一种可迁移性强、训练及运行效率高、样本量标注较少的半监督群体异常行为检测模型。结合卷积自编码器和卷积长短期记忆网络，分别对图像样本进行空间维度特征压缩重构和时间维度运动特征叠加，并通过自动检测实时监控视频图像，及时发现不同场景下人群状态的异常行为，以降低危害乘客财产生命安全的可能性。基于现有的视频监控系统，所提模型可以完成对车站内高发区域和场景的全覆盖，将监控图像每个像素的重构误差以热图或散点图的形式叠加到原始图像中，有助于监管人员迅速发现异常区域，并及时响应处置。在Subway、CHUK Avenue及自建的无锡地铁数据集上验证了所提模型的准确性。研究结果表明，与同类经典模型相比，所提模型在提升检测精度的同时也能基本满足地铁车站场景下的实时性应用需求。
Abstracts:
To solve the public safety risk issues caused by group abnormal behavior in metro such as crowding, running and trampling, targeting the wide variety, poor balance, complex and diverse background environment problems of metro station monitoring video samples, a semi\|supervised group abnormal behavior detection model with strong mobility, high training and operation efficiency, and less sample labeling is proposed. Combined with convolutional auto\|encoder and convolutional long/short term memory network, the image samples are compressed and reconstructed with spatial dimension features, and superimposed with temporal dimension motion features respectively. Meanwhile, real-time monitoring video images are automatically examined, thus group abnormal behaviors in different scenarios are identified in time, lowering the threat on public/private property and safety. Based on the existing video monitoring system, the proposed model can accomplish full coverage of high\|incidence areas and scenes in the station, and the reconstruction error per pixel of the monitoring image is superimposed into the original image in the form of heat map or scatter plot, which is helpful for the supervisors to quickly locate abnormal areas and respond in time. Accuracy of the proposed model is verified by Subway, CHUK Avenue and the self\|built Wuxi Metro dataset. Research results show that compared with similar classical models, the proposed one can elevate detection accuracy while generally meeting the real\|time application requirements in metro station scenario.
过刊浏览
资料下载
友情链接