Object Detection

Object Detection

For the implementation of object detection, a custom-trained YOLOv8 model was employed. It runs object detection on each frame in real-time, classifying each cone and placing bounding boxes around each one. The main goal of this portion of the vision is to locate and classify the cones. This pipeline is integrated within vision_cone_detector, as it is tightly coupled with the rest of the vision algorithms.

Overview

File Structure: vision_cone_detector

• vision_cone_detector/cone_detection.py
• launch/vision_pipeline.launch.py

Tech Stack:

• sensor_msgs
• vision_msgs
• cv2
• cv_bridge
• torch
• ultralytics

Process

The object detection algorithm runs using the following process:

• Subscription to the point cloud on the topic /camera/image/raw

• YOLOv8 is run, objects are detected

• Bounding boxes are rendered on the frames

• This new frame is then published as an image on the topic /camera/image/cones