基于人车交互行为模型的上下客行为识别

上下客行为是常见的人车交互交通行为，但随意地在路边或者禁停区域上下客，不但容易干扰道路交通秩序，还可能造成人员伤亡的恶性交通安全事故，需要及时检出以便疏导管理。受益于智慧灯杆的开发和部署，全路段的上下客行为检测成为可能。设计了一种基于智慧灯杆监控视频的人车交互行为模型HVIB（Human-Vehicle Interaction Behavior）及上下客行为识别方法，实现路边停车和上下客行为的检测。人车交互行为模型HVIB由车辆运动状态检测模块和人车关系检测模块组成。在车辆运动状态检测模块中，利用YOLOv4（You Only Look Once，Version 4）目标检测模型和SORT（Simple Online and Realtime Tracking）跟踪算法输出高置信度目标信息，并抽取车辆时空位置特征表达。在人车关系检测模块中，结合人与车辆的空间位置变化和相对运动方向，形成人车关系的时空特征表达。通过计算视频中人车时空位置特征，基于车辆运动状态判别函数和人车关系判别函数输出车辆运动状态和人车关系类别，并依据不同人车交互行为的定义，可以实现上下客行为识别。使用真实城市交通场景视频数据，对多种天气条件（晴天、阴天、雨天）下的不同人车行为进行了识别试验。试验结果表明：所提出的方法可以全天候工作，其中在白天多种天气条件下，停车和上下客行为的检测准确率能达到90%和87%以上，夜晚正常天气条件下分别为82.5%和77.5%；同时，检测速度在每秒30帧以上，满足实际应用的实时性要求。

Boarding and Alighting Behavior Recognition Based on Human-vehicle Interaction Behavior Model

Boarding and alighting behavior is a typical human-vehicle interaction behavior. However, boarding and alighting passengers randomly on roadsides and in no-parking areas may cause traffic disorder and severe traffic casualties, so it needs to be detected in time to ensure efficient traffic management. Following the development and deployment of smart light poles, the boarding and alighting behavior of the entire road can be detected. In this study, a human-vehicle interaction behavior (HVIB) model was established using smart light pole monitoring videos, and a method of boarding and alighting behavior recognition was developed. The HVIB model consisted of a vehicle-motion state detection module and a human-vehicle relationship detection module. In the vehicle-motion state detection module, the YOLOv4 object detection model and the SORT tracking algorithm were used to obtain high confidence object information and extract vehicle spatiotemporal position features. In the human-vehicle relationship detection module, combined with the spatial position change and movement direction of the pedestrian relative to the vehicle, the spatiotemporal feature expression of human-vehicle relationship was formed. The spatiotemporal position characteristics of humans and vehicles in the video were calculated, and the vehicle motion state and the relationship between humans and vehicles were output based on the vehicle motion state function and the human-vehicle relationship function, respectively. Next, the boarding and alighting behaviors were recognized according to the definitions of different human-vehicle interaction behaviors. Actual data obtained from urban traffic scene videos were used to perform recognition experiments of different human-vehicle behaviors under different weather conditions (such as sunny, cloudy, and rainy). The experimental results show that the proposed method can work day and night. The recognition accuracy of the parking and the boarding and alighting behavior can exceed 90% and 87%, respectively, under various weather conditions during the day. The corresponding values under normal weather conditions at night are 82.5% and 77.5%. The detection speed exceeds 30 frames per second, satisfying the real-time requirements for practical applications.