Design robber detection from surveillance video

You’re a Data Scientist on a team building a computer-vision system for public-safety monitoring. ## Problem Design an ML system that uses fixed surveillance cameras to **detect potential robbers in real time** and trigger an alert to a human operator. ## What to cover 1. **Define the task precisely** - What is the model predicting (e.g., “robbery in progress”, “person-of-interest”, “weapon present”, “suspicious behavior”)? - What is the unit of prediction (frame, clip, track, event) and decision thresholding? 2. **Data & labeling strategy** - Data sources (historical footage, staged data, public datasets) and how you will label positives/negatives. - How you’ll handle **rare events**, noisy labels, and ambiguous cases. 3. **Modeling approach** - Candidate pipelines (detection + tracking + action recognition, multimodal, temporal models, etc.). - How you will deal with occlusion, lighting, camera angle differences, and domain shift. 4. **Evaluation plan** - Choose **primary metric(s)** and explain tradeoffs (e.g., precision/recall, false alarm rate per hour, detection latency). - Offline evaluation vs online evaluation; how you’d estimate performance under class imbalance. 5. **Deployment & monitoring** - Latency/throughput constraints, edge vs cloud inference, model updates. - Monitoring for drift and performance regression. 6. **Risk, fairness, and privacy** - Potential harms (misidentification, demographic bias, privacy violations). - Mitigations, governance, and human-in-the-loop design. Assume you can ask clarifying questions about operational constraints (camera FPS, compute budget, allowable false alarm rate, legal/privacy requirements).