⚠️ WARNING: the development of this project has been migrated here, please reference that repository for the most up-to-date version.

Visual Mapping Pipeline for Autonomous Racing

This repository contains the implementation of a visual mapping pipeline developed for Formula Student Driverless (FSD) autonomous racing competitions. The system performs robust traffic cone detection, real-time vehicle localization, and global mapping using stereo vision and sensor fusion techniques.

🚗 Overview

The pipeline consists of five core components:

1. Cone Detection – YOLOv11-m object detector fine-tuned on the FSOCO dataset.
2. Visual-Inertial SLAM – Real-time vehicle odometry from stereo + IMU.
3. Depth Map Generation – Stereo triangulation for per-pixel depth estimation.
4. Fusion Module – Projects 2D detections into 3D camera and car frames.
5. EKF Mapping – Builds and maintains a persistent global map of cone landmarks.

All modules are implemented as ROS 2 nodes and designed for real-time execution on embedded systems.

🧠 Features

• Accurate cone localization under dynamic racing conditions
• Modular ROS 2 architecture for scalability and integration
• Bird’s-eye-view visualization via image homography
• EKF-based global map refinement and outlier suppression
• Evaluation with both real-world and simulated FSD data

🧪 Dataset

The detector is trained on the FSOCO dataset, which includes over 11,000 annotated images of traffic cones in diverse lighting and track conditions.

🛠️ Setup

Dependencies

• ROS 2 Humble (or compatible distro)
• Python 3.10+
• OpenCV
• NumPy, SciPy
• Ultralytics YOLOv8
• ZED SDK (optional)