Beyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
summary
Computer vision pipeline that counts and tracks road users in Barcelona using YOLOv7 on COCO and a Norfair tracker, deployed on seven RGB cameras recording three days (~6 GB) at 30→15 fps and 640×480 resolution. Pipeline flattens lens distortion, subsamples frames, and outputs per-minute counts; bikes drop accuracy because riders occlude bicycles. Video corruption forces ffmpeg repair in processing. Roadmap moves detection to the camera for edge inference to send counts not raw video.
AI-written summary · about
chapters (12)
- 0:00Introduction to talk
- 1:03What is computer vision
- 3:35Data preprocessing steps
- 4:36Common CV algorithms
- 5:40Urban monitoring project
- 8:50Preprocessing for accuracy
- 9:34YOLO7 and tracking setup
- 10:38Results and visualization
- 12:03Calculating model accuracy
- 13:29Lessons learned and next steps
- 15:59Processing at the edge
- 16:19Q&A session
related talks
47mAn unexpected journey into computer vision, or There and Back Again
35mLarge-Scale Video Intelligence
31mCV at the Dutch Tennis Federation: Utilizing YOLO to create insights for coaches
12mCascade models in computer vision to boost accuracy and performance
21mLeveraging Python for Real-Time Image processing
33mArchitecting Scalable Multi-Modal Video Search
27mAnalysis of sports scenes using computer vision models
30mWhen performance matters: running a real-time computer vision pipeline