Repository contains my bachelor's thesis at ITMO University. The paper was about development of a distributed algorithm for object detection in videos. Used technologies:
- asyncio;
- OpenCV (with Tengine);
- ZeroMQ.
pres.out.mp4
Developed algorithm was tested on a cluster with 22 single-board computers ROCKPro64. To ease deployment process, Docker was installed on each node and then all nodes were joined in Docker Swarm. Complete configs for deployment can be find in the src/deploy directory. A deployment diagram:
Two object detection models were used for testing: yolov4-416 and yolov4-tiny. More information about these models you can find in Darknet repository. Results for yolov4-tiny:
Results for yolov4-416:
Comparison of a distributed solution and processing video on a single node:
| Testing parameters | FPS_1 | W | FPS_W | FPS_W / FPS_1 |
|---|---|---|---|---|
| 1280x720@yolov4-tiny | 1,92 | 4 | 7,64 | 3,98 |
| 1920x1080@yolov4-tiny | 1,71 | 4 | 4,95 | 2,89 |
| 1280x720@yolov4-416 | 0,22 | 21 | 4,22 | 19,18 |
| 1920x1080@yolov4-416 | 0,22 | 21 | 4,14 | 18,82 |
Where:
- FPS_1 - resulting FPS in a single node mode. Use script src/video.py to reproduce the results;
- W - number of worker instances used in a distributed mode;
- FPS_W - resulting FPS in a distributed mode.
To get more detailed results read text version of this work.