Katkat
Back

Design robber detection from surveillance video

Capital OneData Scientist · Technical Screen
EasyData ScientistTechnical ScreenMachine Learning

Last updated

This question assesses the ability to design an end-to-end computer vision pipeline for detecting rare, high-stakes events in real-time video surveillance.

Problem Statement

You are tasked with building a computer vision system for a bank to detect "robbery-in-progress" events from existing surveillance camera feeds. The system must trigger an silent alert to security personnel when a potential robbery occurs.

Requirements

  1. Model Architecture: Propose an approach to detect robbery-like behaviors (e.g., brandishing a weapon, forced entry, or aggressive posturing) in a video stream.
  2. Data Strategy: Robberies are extremely rare events (class imbalance). How would you source, curate, and label a dataset to train this model effectively?
  3. Deployment: The bank requires low-latency inference on existing hardware. How do you ensure the system operates in near real-time without overwhelming on-site compute resources?
  4. Performance Metrics: How will you evaluate this system, considering the trade-offs between precision and recall in a security context?
  5. Ethics: What potential biases or false positive risks exist in this system, and how would you mitigate them?

Expected Output

Provide a high-level design document covering:

  • A data pipeline strategy for rare event detection.
  • The choice of model architecture (e.g., frame-based vs. temporal models).
  • A strategy for minimizing false alarms while maintaining high sensitivity.
  • Ethical safeguards for automated surveillance.

Hints

Hint 1

Consider using a two-stage approach: a lightweight frame-based filter for motion detection followed by a temporal model (like an LSTM or Transformer) for action classification.

Hint 2

Address class imbalance by utilizing synthetic data, video augmentation, or few-shot/anomaly detection techniques instead of just collecting real-world robbery footage.

Hint 3

Think about the impact of false positives—how would a false alarm affect the bank staff and daily operations?

Follow-up questions

  • If your model reports a 98% accuracy but a very high false-positive rate, how would you refine the deployment strategy to improve the utility for security personnel?
  • How would you handle lighting variations, camera angles, or occlusions (e.g., masks/hoodies) that might obscure the perpetrator's actions?
  • What role does 'human-in-the-loop' design play in this system, and how would you incorporate it into your architecture?