-
Track just about anything that detectors can detect
-
Currently its using yolov5 for detections and Deepsort with Siamese Network for tracking
-
Siamese Network is trained on Nvidia AI City Challege Data (640 epochs) and VeRI Wild Dataset (33 epochs) pretrainied weights are provided
-
This Project is built on top of https://github.com/abhyantrika/nanonets_object_tracking and adds detector and capabilities to inference on video/live_feed
-
Currently feature extracter is trained to extract features for vehicals but can be easily trained for other task also
- Download Yolov5 weights from this link https://github.com/ultralytics/yolov5/releases/download/v5.0/yolov5x6.pt
pip install -r requirements.txt
-
python custom_deepsort.py --yolo_weights PATH_TO_DOWNLOADED_WEIGHTS \ --source PATH_2_VIDEO_FILE \ --device GPU_ID_2_USE
- Download dataset and save it in
object_tracking/datasets/train
andobject_tracking/datasets/test
(make sure format of data is correct i.e train/car_id/**images) - change default config
python siamese_train.py
python siamese_test.py
- Beaware this dose not reid the object with same id if its out of frame for long interval of time