yehengchen/Object-Detection-and-Tracking

What it solves

This project tackles the challenge of detecting and tracking objects in images and video, with a focus on real-time performance and integration with robot systems. It brings together popular one-stage detectors (like YOLO and SSD) and two-stage detectors (like R-CNN variants) into a single practical toolkit, so you can train models on custom data and run them on live camera feeds or in ROS-based robots.

How it works

The repository is organized into one-stage and two-stage detection approaches. For one-stage, it provides ready-to-use pipelines combining YOLOv3/YOLOv4 with tracking algorithms (SORT and Deep SORT) for counting pedestrians or vehicles and even measuring social distance. It also includes Darknet_ROS for combining YOLOv3 with ROS for grasp detection, and a DeepLabv3+ ROS version for tracking rocks in a Mars rover simulation. The two-stage section covers Region-based CNNs (Fast R-CNN, Faster R-CNN, Mask R-CNN) with instructions to train Mask R-CNN on your own images. To support custom training, the repo explains how to prepare datasets in COCO or Pascal VOC format and convert them to the YOLO format.

Who it’s for

This is aimed at robotics and computer vision developers who already have some familiarity with deep learning. The setup mentions Ubuntu 16.04/18.04, ROS Kinetic or Melodic, specific NVIDIA GPUs (GTX 1080Ti or RTX 2080Ti), and Python 2.7/3.6. If you want to build object detection and tracking into your own ROS-based robot or need practical example pipelines, this collection gives you step-by-step links and tested configurations.

Highlights

• Combines multiple detection architectures (YOLO, SSD, R-CNN family) in one place
• Real-time tracking options: SORT and Deep SORT for counting and monitoring
• ROS integrations for robot perception, including grasp detection and a Mars rover example
• Guides for training YOLO and Mask R-CNN on custom datasets
• Includes dataset conversion tools (COCO to YOLO, VOC to YOLO)
• Links to educational resources explaining key concepts like mAP and IoU