A web app for monitoring safe distance on Stanford Drone Dataset. Trained on Darknet YOLOv4 for performing small object detection and then performed Tracking by detection on two classes Pedestrians and Bikers using DeepSORT. A violation index is then set(in #pixels i.e. 45 pixels) after several trials and measured the number of total violations in a Stanford drone video. Please checkout my object-detection-on-aerial-videos repo for performing multiobject detection on this dataset. Here a larger dataset is used than before. To get an overview of how it works, please go through my ppt slides(in presentation mode). To get a generalized yolov4-DeepSORT implementation checkout theAIGuysCode-yolov4-deepsort repo. I've done several modifications for my own purpose in this project. Deployed using Flask API. I'll try give a tutorial on how DeepSORT worked here.
 |
|---|
- Train on Darknet yolov4 with larger dataset.
- Apply DeepSORT for multi object tracking.
- Deploy using Flask API.
- Comparative study with other state-of-the-art object detectors.
- To make compatible for edge devices like IOT and mobile devices (Need to train with yolo-v4 tiny model).
- Increase fps (to make realtime).
- Host on cloud.