Project Details: ROS Object Tracking
This project implements a comprehensive ROS2 (Humble) object detection and tracking system within a single package (sim_cam_pkg). The pipeline starts with a simulated camera feed from a video file, performs object detection using a YOLOv11n model (via OpenCV DNN), tracks objects across frames using an OpenCV Kalman filter, and visualizes the results through rqt_image_viewer.
To simulate a continuous camera feed from a single video file, a simple loop detection logic is incorporated. Each time the video loops, the tracker IDs are reset to ensure distinct tracking for each pass.
System Architecture & Pipeline
A key feature includes a loop detection logic for continuous camera feed simulation from a single video, resetting tracker IDs upon each loop.
Demonstration
Key Features:
- Object Detection: Utilizes YOLOv11 for accurate detection.
- Object Tracking: Employs OpenCV’s Kalman filter for smooth tracking.
- Simulated Environment: Works with video files as a simulated camera source.
- Continuous Loop Simulation: Handles video looping and tracker ID resets.
- Visualization: Integrates with
rqt_image_viewer. - Docker file - Base ROS uses OpenCV 4.5
Key Technical Choices:
- Object Detection: YOLOv11n (COCO pre-trained
.onnxmodel) with OpenCV DNN for optimized CPU inference. Achieved ~10-12 FPS on an i5 13th gen CPU. - Tracking Algorithm: Kalman Filter implemented in C++ with a simple greedy approach for track association.
- Build System:
ament_cmaketo support both C++ and Python nodes. - Custom ROS Messages: Defined for structured inter-node communication (e.g.,
DetectionArray,TrackedObjectArray).