OpenTalks.AI - Дмитрий Пагин, Fast cars detection and traffic estimation​

Transcript

1. Fast cars detection and traffic estimation Dmitriy Pagin, ML and CV developer
2. Task Road traffic analysis in Russia is manual. It takes more than 8 hours for 15 minutes video today
3. Task • detect cars
4. Task • detect cars • track cars
5. Baseline - people tracking
6. Problems Cars: - faster (2 metres per frame!) - smaller (10 px in minimal dimension) + more predictable movement
7. YOLOv2 - blinking - problems on small cars - problems on edges
8. YOLOv2 1 fps
9. YOLOv3 - bigger + accurate on small + fullHD frame + robust
10. YOLOv3 7 fps
11. > 70k cars on 4k images Dataset
12. better than 1024x1024x1 Learning and Fine-tuning - 608x608 px - batchSize = 3 - custom augmenters
14. Learning and Fine-tuning - 608x608 px - batchSize = 3 - custom augmenters - Radam optimizer (instead warmup + reduce LR) - Hard negative mining for trucks
15. Learning and Fine-tuning - 608x608 px - batchSize = 3 - custom augmenters - Radam optimizer (instead warmup + reduce LR) - Hard negative mining for trucks mAP75 = 0.96
16. Baseline Inference Speed 7 fps
17. Weights Pruning
18. Weights Pruning -25% convs = size: 240 mb mAp: 0.9656 inf: 150 ms size: 155 mb mAp: 0.9622 inf: 100 ms 10 fps
19. OpticalFlow step or classical cv is alive ! - find good features to track - calculate sparse optical flow
20. OpticalFlow step 19 fps Calculation doesnt work for 3 consistent frames
21. Speed extrapolation step - estimate speed as pixels/frame - extrapolate next position 28 fps
22. Final pipeline 1 2 3 4 5 6 Update trajectories 4 5 6 step 1 step 2 Speed Extrapolation OpticalFlow YOLOv3 Detection Engine
23. 1 fps -> 28 fps on FULLHD
24. Tracking - IoU - Color descriptor (it’s enough!)
25. Bridges! - Allowed zone by motion vector - Size overlap - Color descriptor
26. Bridges! - Allowed zone by motion vector - Size overlap - Color descriptor
27. Thanks! Questions? dm.pagin@gmail.com +7 952 335 65 70
28. Appendix. Examples
29. Appendix. Examples
30. Appendix. Examples
31. Appendix. Yolov3
32. Weights Pruning Шаг mAP75 Число параметров, млн Размер сети, мб От изначальной, % Время прогона, мс Условие обрезания 0 0.965 60 241 100 150 - 1 0.962 55 218 91 140 5% от всех 2 0.962 50 197 83 132 5% от всех 3 0.963 39 155 64 112 15% для слоев с 400+ сверток 4 0.955 31 124 51 100 10% для слоев с 100+ сверток
33. Appendix. Radam
34. Pruning convs
35. Pruning convs. Good choice 2000
36. Pruning convs. Bad choice 25
37. Pruning flat