面向数据驱动建模的行人轨迹自适应压缩与误匹配识别分割

Research on Adaptive Trajectory Compression and Mismatch Trajectory Identification-segmentation Method for Data Driven Modeling

轨迹数据驱动的行人行为分析建模在公共场合异常事件监测、人车冲突风险评估等方面具有重要意义,广布的交通视频监控是行人群轨迹数据的重要来源。行人轨迹具有趋势性和规律性,提取的原始轨迹信息冗余较大,且密集行人群频繁遮挡,不同行人轨迹易发生误匹配,导致数据失真。针对以上问题,根据行人轨迹的局部结构特征和数值特性,设计一种改进的两阶段自适应滑窗轨迹压缩算法ATSSW(Adaptive Two Stage Sliding Window)和基于轨迹局部转向角的误匹配识别和分割方法ABTDS(Angle-based Trajectory Detection and Segmentation),清洗和压缩行人轨迹数据。首先,ATSSW算法考虑轨迹各坐标分量的数值分布特征,将提取到的所有原始轨迹分为漂移和非漂移2类,采取不同的策略分别压缩2类轨迹;然后,ABTDS算法分析压缩后的轨迹局部转角特征,辨识误匹配轨迹样本;最后,ABTDS算法分割误匹配样本,并用分割后的轨迹更新原始轨迹数据集。研究结果表明:ATSSW算法压缩了653条原始行人轨迹,总压缩信息损失1 002.04,总平均轨迹压缩率为6.07%,总平均轨迹压缩保留率为95.35%;原始轨迹集中存在126条误匹配轨迹,ABTDS算法辨识并成功分割了其中的107条,检出率为84.92%;所提算法抑制了原始行人轨迹中漂移点和误匹配现象所致的干扰,减少了原始轨迹数据噪声,可提高轨迹数据驱动的行人行为建模精确度;适当压缩原始轨迹,可减轻轨迹数据存储处理的负担。

Data-driven pedestrian trajectory analysis and modeling are critical in abnormal detection and vehicle-pedestrian collision avoidance. As video surveillance is now widely distributed in cities, it has become a reliable source of pedestrian trajectory data. However, pedestrian trajectory extraction inevitably suffers from information redundancy because of trajectory regularity, the tendency of crowds to self-organize, and coherence in movement of individual pedestrians. Mismatch problems that are caused by the frequent interaction and occlusion amongst pedestrians exist. To address these problems, an improved adaptive sliding window compression method known as an adaptive two-stage sliding window(ATSSW) and a mismatched trajectory detection method known as angle-based trajectory detection and segmentation(ABTDS) were proposed in this study. Trajectory data were then processed using these two algorithms. Regarding compression, the ABTDS first divided the trajectories into two parts based on the coordinate correlation characteristics of the trajectories and then implemented the compression under different strategies. The mismatched trajectory detection method considered the trajectory corner as a local structural feature, which is an indicator of whether a mismatch exists and where to split the trajectory into parts. Finally, the ABTDS algorithm manipulated and updated the original trajectory dataset. Experimental results show that the ATSSW algorithm compresses 653 pedestrian trajectories, where the average compression error is 1 002.04 and the average compression rate is 6.07% with a preservation rate of 95.35%. With ABTDS, 107 out of 126 mismatched trajectories are found, where the detection rate reaches 84.92%. The proposed algorithms can remove noise from pedestrian trajectory data and can better support data-driven trajectory prediction and modeling.