Lightweight LiDAR-Based Cone Detection Framework Using Random Forest for Formula Student Driverless

Formula Student Driverless (FSD) competitions require vehicles to navigate tracks autonomously with high precision. Many existing perception systems rely on complex deep learning models and multi-sensor fusion, which often necessitate power-hungry GPU hardware. This paper introduces a streamlined, LiDAR-only perception pipeline designed specifically to run efficiently on standard CPU hardware, providing a practical solution for teams working with limited computational resources.

A Streamlined Perception Pipeline

To achieve low-latency performance without specialized hardware, the authors developed a multi-stage processing pipeline. The system begins by removing ground points from the LiDAR data and applying IMU-based motion compensation to account for the vehicle's movement. Once the data is cleaned, the system uses DBSCAN clustering to group points into potential objects. Finally, a Random Forest classifier evaluates these clusters based on geometric features to determine if they are traffic cones. The same ai safety question is explored in A Unified Physics-Aware Quantum Machine Learning..., which adds a research perspective.

Optimizing for Efficiency

A key contribution of this research is the optimization of the classification model. By conducting a feature importance analysis, the researchers were able to reduce the number of input features required for the Random Forest model from 12 down to 7. This reduction was achieved without sacrificing the accuracy of the system, allowing for a more lightweight model that maintains high performance while reducing the computational load.

Performance and Real-World Results

The framework was evaluated using a dataset of 2,371 labeled clusters gathered from actual FSD racing events. The results demonstrate that the pipeline is highly effective, achieving an F1-score of 98.33%. Furthermore, the system is exceptionally fast, boasting an end-to-end runtime of just 3.13 milliseconds on CPU-only hardware. The same computer vision question is explored in CUA-Universe, which adds a research perspective.

A Resource for the Racing Community

Beyond the technical framework, the authors have released their dataset, labeling tool, and trained models to the public. By providing these resources, the researchers aim to establish a reproducible baseline for other autonomous racing teams. This contribution is intended to help teams with constrained hardware budgets implement reliable, high-speed perception systems for their own driverless vehicles. The same ai safety question is explored in SIRF, which adds a research perspective. as detailed in the full paper on Arxiv