SENSE-VAD
收藏资源简介:
SENSE-VAD是由南佛罗里达大学团队构建的首个面向自动驾驶的合成视频异常检测基准数据集,专注于社会复杂性异常场景。该数据集包含108个标注视频片段,总计约540,888帧,数据通过CARLA仿真平台与Unreal Engine在多版本引擎下生成,涵盖儿童安全、群体行为、人-物交互等6大类15个异常类别,并辅以真实世界视频片段以支持仿真到现实的迁移验证。数据创建过程采用参数化脚本控制,严格分离社会关系异常与运动统计异常,通过多样化场景配置避免模型过拟合仿真器视觉特征。该数据集旨在解决自动驾驶系统中对行人社会语境异常(如儿童远离监护人、群体追逐等)的检测盲区,推动感知系统在复杂社会交互场景下的安全预警能力。
SENSE-VAD is the first synthetic video anomaly detection benchmark dataset tailored for autonomous driving, developed by a team from the University of South Florida, focusing on socially complex anomalous scenarios. This dataset includes 108 annotated video clips, totaling approximately 540,888 frames. The data is generated across multiple versions of the CARLA simulation platform and Unreal Engine, covering 15 anomalous categories under 6 major groups such as child safety, group behavior, human-object interaction and more. Real-world video clips are also supplemented to support sim-to-real transfer validation. The data creation process is controlled by parametric scripts, which strictly separate social relationship anomalies and motion statistical anomalies. Diverse scene configurations are adopted to prevent models from overfitting to the visual features of the simulator. This dataset aims to address the detection blind spots of pedestrian social context anomalies (e.g., children wandering away from guardians, group chasing, etc.) in autonomous driving systems, and promote the safety early warning capability of perception systems in complex social interaction scenarios.
数据集概述
- 名称: Anonymous Dataset Submission for peer review
- 描述: 用于同行评审的匿名上传数据集,服务于研究论文 "SENSE-VAD: Sentient and Semantic Video Anomaly Detection for Autonomous Driving"(基于感知与语义的视频异常检测在自动驾驶中的应用)。
- 资源类型: Dataset(数据集)
- 发布者: Zenodo
- DOI: 10.5281/zenodo.20955310
- 版本: v2(当前版本)
- 发布日期: 2026年6月27日
- 许可证: Apache License 2.0
文件信息
- 文件数量: 1个文件
- 总文件大小: 22.3 GB
- 包含文件:
- SENSE_VAD_Dataset_Annotation_Generation_Code.zip
- 大小: 22.3 GB
- MD5校验码: 514fcbe780485f907ceb975bc8b35acd
- SENSE_VAD_Dataset_Annotation_Generation_Code.zip
其他信息
- 可用日期: 2026年6月26日
- 历史版本:
- v2(当前版本): DOI 10.5281/zenodo.20955310
- v1: DOI 10.5281/zenodo.20947278
- 引用所有版本DOI: 10.5281/zenodo.20947277

- 1SENSE-VAD: Sentient and Semantic Video Anomaly Detection for Autonomous Driving南佛罗里达大学 · 2026年



