How We Built a Safety Violations Detection System
summary
A safety-violation detection system for video streams combines a modified Faster R-CNN detector with object tracking and Kalman interpolation so the detector runs at about 6 FPS while the overall stream meets real-time requirements. The team trained on ~200 hours (~40k frames), achieved mAP > 0.9 on internal data, implemented the pipeline with PyTorch/Catalyst and exported with TorchScript/ONNX, and delivered a prototype in ~1 month with production readiness in 3–4 months plus a dashboard and per-shift reports.
AI-written summary · about
chapters (11)
- 0:00Violation detection intro
- 2:51Product evolution
- 6:06Switching to trackers
- 8:55App and visualization
- 12:30Framework details
- 17:51Data labeling
- 18:12Model optimization
- 20:22Metrics and NN interpretation
- 25:36Production optimization
- 28:44Development plans
- 31:12Q&A
related talks
35mFootball video analysis using Deep Learning
34mTesting Live Video Streams At Scale
41mPractical Aspects of Applying Modern Computer Vision Technologies
30mCrash detector
53mVideo Object Detection at Scale
33mUsing machine learning to improve player safety in the NFL
27mAnalysis of sports scenes using computer vision models
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark