Source-linked AI summary
Lightweight LiDAR-Based Cone Detection Framework Using Random Forest for Formula Student Driverless
Márk Mező-Kerekes, Péter Praksz, Chang Liu
TL;DR
Formula Student Driverless perception must meet strict computational and latency constraints while maintaining accurate cone detection. This paper presents a lightweight LiDAR-based pipeline for CPU-only operation and reports 98.33% F1-score with a 3.13 ms runtime.
Problem
FSD perception must handle strict computational budgets, vehicle motion, accurate cone detection, and limited annotated data, while GPU-dependent solutions can be infeasible for many student teams.
Method
The paper presents a LiDAR-based perception pipeline designed to operate efficiently on CPU systems without GPU acceleration, using importance-based feature reduction while maintaining classification performance.
Results
98.33% F1-score and 3.13 ms CPU-only runtime were achieved for the lightweight LiDAR-based cone detection pipeline.
Takeaways & Limitations
The compact geometric-feature and Random Forest approach is suitable for resource-constrained real-time perception, with approximately 121MB average RAM usage and released tools and data supporting deployment and reproduction.
Takeaways & Limitations
Beyond 15m, sparse cone clusters increase missed detections or poorly conditioned feature estimates, and the system has not been evaluated in rain or fog.
Abstract
from arXiv · showhide
Reliable, low-latency perception is crucial for Formula Student Driverless vehicles, yet many existing pipelines rely on deep learning and multi-sensor fusion, often requiring GPU acceleration. This paper presents a lightweight LiDAR-only perception pipeline tailored for CPU execution, combining ground removal, IMU-based motion compensation, DBSCAN clustering, and geometric feature-based Random Forest classification. Feature importance analysis reduced the model input from 12 to 7 features while preserving performance. Evaluated on 2,371 labeled clusters collected from real FSD events, the pipeline achieves an F1-score of 98.33% and an end-to-end runtime of 3.13 ms on CPU-only hardware. The released dataset, labeling tool, and trained models provide a practical and reproducible baseline for other resource-constrained autonomous racing teams.
1 Introduction
Formula Student Driverless requires accurate cone detection under tight computational and hardware constraints. This paper proposes a CPU-efficient LiDAR-based pipeline and contributes feature reduction, annotation resources, labeled data, and reproducible models.
- The pipeline targets FSD perception challenges involving strict computational budgets, dynamic vehicle motion, limited annotated data, and infeasible GPU-dependent solutions for many teams.
- The proposed LiDAR-based perception pipeline is designed to operate efficiently on CPU-based systems without GPU acceleration.
- Importance-based feature reduction improves computational efficiency while maintaining classification performance.
- The paper releases a public labeling and training tool for LiDAR cone clusters and training models on labeled data.
- The study provides 2,371 manually annotated clusters from real FSD events, including yellow, blue, and orange cones.
- The released trained models are intended as reproducible baselines for future FSD teams.
2 Related Work
Prior FSD perception methods include sensor fusion, LiDAR-only deep learning, rule-based clustering, and traditional machine learning. Lightweight LiDAR-only CPU methods remain underexplored despite the efficiency advantages of handcrafted features and classifiers.
- Multi-sensor fusion can provide reliable perception but requires substantial computation, hardware, and annotated data, limiting resource-constrained applicability.
- LiDAR-only deep learning reduces sensor complexity but still depends on large datasets and GPU acceleration.
- Rule-based Euclidean or DBSCAN pipelines have demonstrated viable FSD results but require careful manual tuning.
- Handcrafted geometric features with SVMs or Random Forests offer a lightweight CPU-based efficiency-accuracy trade-off, yet remain underexplored for FSD.
3 Methodology
The pipeline preprocesses 20 Hz LiDAR scans through ground removal, IMU-based motion compensation, and DBSCAN clustering before classifying clusters with a feature-based Random Forest. Feature selection reduces the input from 12 to 7 geometric features while preserving performance and improving efficiency.
- 3 Methodology: The system processes 20 Hz scans from an Ouster OS1-32 sensor through preprocessing and Random Forest classification stages.The dataset contains clustered LiDAR segments from Skidpad, Acceleration, and Autocross scenarios, with 2,371 manually annotated samples and run-level separation for train/test splits.
- 3.2 Preprocessing: The preprocessing stage removes ground points using sector-based radial and height thresholds, then applies IMU-based motion compensation and DBSCAN clustering.Ground removal reduces computational overhead, motion compensation addresses scan distortion, and DBSCAN groups dense regions while handling noise without a predefined cluster count.
- 3.3 Feature Extraction and Classification: Seven selected geometric features feed a Random Forest that classifies each cluster as cone or non-cone after redundant features below 1% importance are removed.The reduction from 12 to 7 features improves computational efficiency while preserving and slightly improving classification performance.
- 3.3 Feature Extraction and Classification: Width and depth are assigned dynamically from the larger and smaller x-y spreads, respectively, producing approximately a 1% F1-score increase across viewpoints.This normalization is intended to improve robustness to orientation variations.
- 3.3 Feature Extraction and Classification: The Random Forest uses 65 estimators and maximum depth 10, with hyperparameters selected by grid search and 5-fold cross-validation.A fixed random seed of 42 supports reproducibility.
4 Results
The pipeline achieves high cone-detection performance with low CPU runtime and compares favorably with published FSD systems, while direct runtime comparisons remain approximate because hardware differs.
- The evaluation used a CPU-only Intel Core Ultra 9 285H system with 32GB RAM running Ubuntu 24.04 in performance mode.
- The Random Forest classifier achieves an F1-score of 98.33% on the held-out test evaluation.The evaluation used 504 test samples.
- 3.13 ms end-to-end runtime corresponds to approximately 317 Hz, substantially exceeding the 20 Hz LiDAR scan rate.
- 98.33% F1-score is maintained up to 15 m while the pipeline runs faster than the compared systems without GPU acceleration.The paper notes that competing runtime measurements use different hardware, so comparisons are inherently approximate.
5 Discussion, Limitations and Conclusion
The discussion presents the reduced-feature pipeline as efficient and suitable for constrained hardware, while identifying distance and weather conditions that limit validated deployment scope.
- 3.13 ms runtime and stable precision and recall after feature reduction indicate that seven features preserve performance while reducing computational cost.The paper attributes the removed features’ redundancy to the consistent runtime improvement shown in Table 3.
- Approximately 121MB average RAM usage supports deployment on resource-constrained hardware without modification.
- Random Forest class probabilities can support uncertainty-aware decisions in downstream mapping and path-planning modules.
- Beyond 15 m, sparse OS1-32 returns increase missed detections or poorly conditioned feature estimates, and adverse weather has not been evaluated.The stated weather boundary includes rain and fog.
- The conclusion reports 98.33% F1-score and 3.13 ms CPU-only runtime, with future work targeting color detection and adverse-weather robustness.