Here's our project
Part of the Freshman Research Initiative Autonomous Robots Stream, we were required to research a concept related to robotics. This project was developed in a group of 3: Taraka Pranav Rayudu, Ajith Kemisetti, and myself. Our models were trained on the provided COCO dataset, and tested on the open-sourced MOT datasets. We used the Darknet YoloV3 and Deepsort Algorithms to first retrieve the bounding boxes of a person in a frame, and track it using the Deepsort. In the background, YoloV3 and Deepsort use Neural Networks to perform object recognition and obtain the bounding boxes. Considering that this was a 3 week long project, we had 58.79% success rate in terms of accurately identifying the people in the data set.The future steps include to increase the classification percentage, move away from either YOLO or Deepsort to focus the identification from one tool itself, and better handle the occlusion of people walking past each other.