SpatioTemporal Graph Dataset for Exposing Perspective and Occlusion Faults in Video Surveillance

This repository hosts a highly controlled, 18-scenario video anomaly detection benchmarking dataset comprising 7,400 high-definition frames designed to isolate and evaluate the physical and geometric failure boundaries of 2D spatio-temporal graph neural networks. While contemporary skeleton-based surveillance tracking models operate as privacy-preserving solutions by tokenizing human actions into abstract coordinate graphs, standard evaluation databases are heavily bottlenecked by uncalibrated web-scraped feeds. Shifting camera pitches, dynamic auto-exposures, and unstable frame rates introduce compounding spatial variables that mask the underlying algorithmic limitations of coordinate-based architectures. To mathematically isolate these elements, this dataset was engineered within a standardized 12 by 8 meter laboratory footprint under uniform, static fluorescent lighting. Visual sequences were recorded utilizing a fixed-position, ceiling-mounted RGB surveillance sensor locked at a resolution of 1920 by 1080 pixels at exactly 30.0 frames per second. The tracking canvas is marked with metric ground-truth boundaries extending along the primary optical axis from Z = 0.5 meters to Z = 6.0 meters, allowing direct quantification of perspective coordinate degradation.本代码仓库托管了一套高度可控的18场景视频异常检测（video anomaly detection）基准测试数据集，包含7400帧高清画面，旨在隔离并评估二维时空图神经网络（2D spatio-temporal graph neural networks）的物理与几何失效边界。 当前基于骨骼的监控跟踪模型通过将人体动作Token化（tokenizing）为抽象坐标图，成为隐私保护型解决方案，但标准评估数据集普遍受限于未校准的网络爬取素材。相机俯仰角变化、动态自动曝光以及不稳定帧率会引入复合空间变量，掩盖基于坐标架构的底层算法局限性。 为从数学层面隔离上述干扰因素，本数据集在标准化的12×8米实验室场地中构建，采用均匀恒定的荧光照明。视觉序列通过固定安装于天花板的RGB监控传感器录制，传感器分辨率固定为1920×1080像素，帧率严格为30.0帧每秒。跟踪视野内沿主光轴标注了公制真值边界，范围为Z=0.5米至Z=6.0米，可直接量化透视坐标退化程度。