Multi Object Tracking

• 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

Demo Video

Steps to run tracker on custom Video

• 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 
  

Training steps

• Download dataset and save it in object_tracking/datasets/train and object_tracking/datasets/test (make sure format of data is correct i.e train/car_id/**images)
• change default config
• python siamese_train.py

Get Test Scores

• python siamese_test.py

Limitations

• Beaware this dose not reid the object with same id if its out of frame for long interval of time

Refrences

• Siamese Net https://www.cs.cmu.edu/~rsalakhu/papers/oneshot1.pdf
• DeepSort Paper https://arxiv.org/abs/1703.07402
• https://github.com/ultralytics/yolov5
• https://github.com/nwojke/cosine_metric_learning/
• https://github.com/nwojke/deep_sort
• https://github.com/abhyantrika/nanonets_object_tracking