基于改进时间金字塔网络的出租车乘客上下车动作识别模型  

A Recognition Model for Passenger Boarding and Alighting Action Based on Improved Temporal Pyramid Network

作　　者：廖惠敏 罗静茗 张璟辉 刘文平 董婉青 肖晖 黄坚[2] LIAO Huimin;LUO Jingming;ZHANG Jinghui;LIU Wenping;DONG Wanqing;XIAO Hui;HUANG Jian(Transportation Comprehensive Enforcement Corps,Beijing 100044,China;School of software,Beihang University,Beijing 100089,China;Zhonglu High tech Transportation Technology Group Co.,Ltd,Beijing 100089,China)

机构地区：[1]北京市交通运输综合执法总队,北京100044 [2]北京航空航天大学软件学院,北京100089 [3]中路高科交通科技集团有限公司,北京100089

出　　处：《交通信息与安全》2024年第6期95-102,共8页Journal of Transport Information and Safety

基　　金：国家重点研发计划项目(2022YFB2602104);北京市交通行业科技项目(0686-2241B1251414Z);车路一体智能交通全国重点实验室自主研究项目(2021-Z011)资助。

摘　　要：传统基于图像处理的违法载客识别算法依赖人工制定的人车交互规则以确定上下车行为的发生。然而,由于交通场景的复杂性,人工制定的规则集不够完善,导致算法识别效果较差。因此,引入基于时间金字塔网络(temporal pyramid network,TPN)的深度学习模型进行上下车动作识别,通过大量样本集的训练提取较为完备的出租车乘客上下车行为特征,提升识别准确性。针对TPN模型无法区别司乘角色身份的问题,重新设计基于车门区域感知的模型输出层,增强模型多维度特征提取效率;针对上下车行为时空跨度大,模型易受无关动作干扰问题,加入一种基于动态窗口权重的滑窗机制,捕捉动作关键视频帧,提高识别效率。综合上述改进措施,提出了基于车门区域感知和动态权重的出租车乘客上下车动作识别模型(boarding and alighting neural network,BANN),实现高效准确的违法载客行为识别。基于首都机场监控视频构建包含4047段带标注视频的训练集和810段未标注视频的测试集对模型进行验证。实验结果表明:BANN模型的查准率和查全率分别达到90.21%和88.53%,较基准TPN模型分别提升了9.78%和11.04%,能够较好满足枢纽场站交通秩序监管的需要。Traditional algorithms for identifying illegal passenger-carrying behavior,which rely on image process-ing techniques,utilize manually crafted human-vehicle interaction rules to discern boarding and alighting actions.However,these rule sets often fall short due to the intricate nature of traffic scenarios,resulting in suboptimal recog-nition performance.Therefore,a deep learning model based on a temporal pyramid network(TPN)is introduced for boarding and alighting action recognition.By training on a large dataset,more complete features of taxi passenger boarding and alighting behaviors are extracted to improve recognition accuracy.To address the issue of the TPN model not distinguishing between driver and passenger roles,the output layer is redesigned based on door area per-ception.This modification enhances the efficiency of multi-dimensional feature extraction.To tackle the issue of the large spatiotemporal span in boarding and alighting actions,which leads to interference from irrelevant move-ments,a sliding window mechanism is introduced.This mechanism,based on dynamic window weights,captures key video frames of the actions,enhancing recognition efficiency.Based on the above improvement measures,a boarding and alighting neural network(BANN)model,based on door area perception and dynamic weights,is pro-posed to efficiently and accurately recognize illegal passenger-carrying behaviors.A training dataset with 4,047 an-notated video clips and a test dataset with 810 unannotated video clips are constructed for model performance vali-dation based on surveillance videos from Beijing Capital Airport.Experimental results demonstrate that the BANN model achieves precision and recall rates of 90.21%and 88.53%,respectively,representing improvements of 9.78%and 11.04%over the baseline TPN model.These results indicate that the BANN model can effectively meet the needs of traffic order supervision in transportation hubs.

关 键 词：智能交通 上下车动作识别 乘客上下车动作识别网络 时间金字塔网络 违法载客 深度学习 

分 类 号：U495[交通运输工程—交通运输规划与管理]