The main files are inside yolov3 folders and see the text file commands to run the program. Also see the pdf file: traffic_light_detection.pdf
Task: Real time traffic light detection. Save the results in json file. Format:
frame_id”: [
“traffic_light_id”: {
“coords”: [x1, y1, x2, y2],
“state”: “red|yellow|green”,
affect”: true|false }
We used bosch small traffic lights dataset train dataset + yellow traffic lights from test set(there was class disblance). Moreover, we have decided to annote the frames (every 20's frame) of the videos, that the lecturer provived us.
We have used frames of 2 videos for validation.
Since the task was real time detection, we have decided to use Yolo, but big yolo wasn't fast enough in our GPUs (GTX 1070), so we used tiny yolo with 23 layers.
We integrated SORT tracker from this github, which uses Kalman Filter to track the objects. It helped to smoothen the results from detector.
See the result with tracker
Aslan Ubingazhibov - Data Science, Higher School of Economics, aubinagzhibov@edu.hse.ru
Alisa Kugusheva - Data Science, Higher School of Economics, akugusheva@edu.hse.ru
https://github.com/ultralytics/yolov3
https://github.com/abewley/sort