Video Anomaly Detection with Seven Human Behavior Classes from Two Camera Views

510 video clips of 10-15 seconds each depict seven classes of unusual human movements relevant to surveillance. The dataset includes 35-40 balanced samples per class, with voluntary participants acting in each class to provide diversity. Zhalgas, Aidana contributed this dataset to Harvard Dataverse, last updated on 2026-06-23.

Use Cases

• Training anomaly detection models based on unusual human movements like seizures or heart attacks.
• Evaluating multi-view video classification algorithms based on the two camera viewpoints.
• Developing behavior recognition systems for surveillance based on classes like 'limp' or 'beating with bag'.
• Benchmarking video-based action recognition models on a balanced set of normal and abnormal activities.

Strengths

• 510 video clips provide a substantial base for model training.
• Balanced distribution with 35-40 samples per class across seven behavior categories.
• Includes two distinct camera viewpoints (camera 16 and camera 17) for multi-perspective analysis.
• Voluntary participants appear in each class, likely providing diversity in appearance and motion patterns.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.
• Last updated 2026-06-23 17:03:08; freshness should be verified.

Provenance

Source

Harvard Dataverse

Collection Method

Voluntary participants acted as actors in the video clips.

Freshness

2026-06-23