EgoPolice: A Benchmark for Egocentric Video Understanding in High-Stakes Police Body-Worn Camera Footage

2026-02-09 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Expert, extended

Summary

EgoPolice is a new benchmark dataset comprising over 180 hours of real, egocentric police body-worn camera (BWC) footage, curated from publicly available videos of police-civilian interactions. It features second-by-second annotations for nine critical action labels, including "Weapon Out" and "Physical Interaction." The dataset presents significant challenges for video understanding models due to rapid camera motion, dense human interactions, and rare high-stakes events, which lead to visual degradation and limited global appearance cues. Benchmarking state-of-the-art models, including VideoMAE V2, X-CLIP, and Gemini 2.5 Pro, revealed that while video-pretrained models perform better in classification, even Gemini 2.5 Pro achieved only 76.9% accuracy on 1-minute clips and struggles with high-risk actions, indicating unsuitability for autonomous deployment. EgoPolice aims to provide a foundation for human-in-the-loop systems to efficiently identify events in vast BWC archives.

Key takeaway

For AI Scientists and Computer Vision Engineers developing systems for sensitive, real-world environments like law enforcement, recognize that current video models, including Gemini 2.5 Pro, are not yet reliable enough for autonomous deployment in high-stakes body-worn camera analysis. You should prioritize human-in-the-loop workflows where AI assists human reviewers by surfacing potentially critical video segments for verification. This approach mitigates the risks of unexamined bias and system failures, enabling scalable auditing and retrospective analysis while preserving human judgment.

Key insights

Egocentric police BWC footage presents unique, high-stakes challenges for video understanding models, requiring specialized benchmarks and human oversight.

Principles

• BWC footage demands motion-robust, context-aware egocentric perception.
• Objective, intent-free action definitions reduce annotation subjectivity.
• Model reliability in high-stakes domains requires rigorous, domain-specific benchmarks.

Method

A two-stage annotation pipeline filters interactions, then labels 9 objective action categories at per-second granularity, using a custom web tool and mental health safeguards.

In practice

• Use EgoPolice to benchmark video models in high-stakes egocentric settings.
• Deploy models in human-in-the-loop systems for BWC footage review.
• Prioritize video pretraining for robust BWC action recognition.

Topics

• Egocentric Video
• Body-Worn Cameras
• Video Understanding
• Dataset Benchmarking
• Action Recognition
• Human-in-the-Loop AI

Code references

• JaidedAI/EasyOCR

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• EgoEverything: A Benchmark for Human Behavior Inspired Long Context Egocentric Video Understanding in AR Environment — EgoEverything uses human attention signals to create a realistic benchmark for egocentric video understanding in AR.
• Divide, Deliberate, Decide: A Multi-Agent Framework for Fine-Grained Egocentric Action Recognition — Multi-agent VLM deliberation with heterogeneous specialists improves fine-grained egocentric action recognition by decorrelating model priors.
• Whole-Body Conditioned Egocentric Video Prediction — PEVA predicts egocentric video from whole-body human actions, enabling embodied planning and long-horizon simulation.
• SkillSpotter: Pose-Aware Multi-View Skilled Action Detection and Grading in Ego-Exo Videos — SkillSpotter jointly detects and grades skilled actions in ego-exo videos by fusing visual and pose data across multiple views.
• HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining — Carefully processed egocentric human video can surprisingly outperform real-robot data for embodied foundation model pretraining, offering a scalable alternative.
• Ego2World: Compiling Egocentric Cooking Videos into Executable Worlds for Belief-State Planning — Ego2World is a benchmark for embodied agents to plan under partial observation using egocentric video-derived executable worlds.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.