UCSD 异常数据集

UCSD 异常检测数据集通过高处俯瞰人行道的固定相机获取，人行道人群密度从稀疏到拥挤不等，正常视频仅含行人，异常事件源于人行道中非行人实体流通或异常行人运动模式（如骑自行车者、滑冰者等自然发生的异常，非人为暂存）；数据分为 Peds1 和 Peds2 两个场景子集，Peds1 含 34 个训练、36 个测试视频样本（人群走向 / 远离摄像机，有透视失真），Peds2 含 16 个训练、12 个测试视频样本（行人运动与摄像机平面平行），视频片段均被分成约 200 帧的剪辑；每帧有指示是否存在异常的二进制标志，Peds1 的 10 个剪辑和 Peds2 的 12 个剪辑还提供像素级二进制掩码，用于识别异常区域

The UCSD Anomaly Detection Dataset is acquired using fixed cameras mounted high above sidewalks, with crowd densities spanning from sparse to congested. Normal videos exclusively feature pedestrians, while abnormal events arise from either non-pedestrian entities circulating on sidewalks or anomalous pedestrian motion patterns—such as cyclists, skaters and other naturally occurring anomalies, excluding those artificially staged. The dataset is split into two scenario subsets: Peds1 and Peds2. Peds1 contains 34 training and 36 test video samples, where pedestrians move either toward or away from the camera with perspective distortion. Peds2 includes 16 training and 12 test video samples, where pedestrian motions are parallel to the camera's imaging plane. All video sequences are divided into clips of approximately 200 frames each. Each frame has a binary flag indicating whether an anomaly is present. Additionally, 10 clips from Peds1 and 12 clips from Peds2 provide pixel-level binary masks for localizing abnormal regions.

搜集汇总

背景与挑战

背景概述

UCSD异常数据集是一个用于行人异常检测的视频数据集，包含Peds1和Peds2两个子集，分别有不同数量和视角的视频样本。数据集提供了二进制标志和部分像素级掩码，用于识别异常事件和区域。