Lightweight LiDAR-Based Cone Detection Framework Using Random Forest for Formula Student Driverless
Márk Mező-Kerekes, Péter Praksz, Chang Liu
cs.AI
Quick verdict
Claims 98.33% F1 at 3.13 ms CPU-only for LiDAR cone detection via Random Forest; solid engineering, but small single-sensor dataset, no same-data baselines.
Summary
This paper presents a CPU-only, LiDAR-only cone detection pipeline for Formula Student Driverless racing, combining ray-based ground removal, IMU-based motion compensation, DBSCAN clustering, and a Random Forest classifier over seven geometric features (height, width, depth, aspect ratio, volume, density, distance from sensor). Trained and evaluated on a custom dataset of 2,371 manually labeled clusters from real FSD events (1,867 train / 504 test, with clusters from the same run never split across sets), the model reaches 99.16% precision, 97.51% recall, and 98.33% F1 on the held-out test split. Importance-based selection reduced an initial 12-feature set to 7, cutting end-to-end runtime from 4.54 ms (9 features) to 3.13 ms (7 features) on an Intel Core Ultra 9 285H while F1 improved slightly from 97.77% (12 features). The authors report ~317 Hz theoretical throughput, ~121 MB RAM, and detection up to 15 m; beyond that range the OS1-32 returns fewer than five points per cone. The dataset, an open-source labeling tool, and trained models are released. Limitations include the 15 m horizon, untested adverse weather, and a single sensor platform; the evaluation also lacks confidence intervals or variance estimates, and no alternative classifier or deep-learning baseline is run on the same data, leaving the competitiveness-with-deep-learning claim unsupported head-to-head.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 18 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A LiDAR-only pipeline (ray-based ground removal, IMU motion compensation, DBSCAN, and a 7-feature Random Forest) achieves 98.33% F1 (99.16% precision, 97.51% recall) on a 504-cluster held-out test set from real FSD events.
- Importance-based feature reduction from 12 to 7 features preserves and slightly improves F1 (97.77% to 98.33%) while cutting end-to-end runtime to 3.13 ms.
- The full pipeline runs on CPU-only hardware with about 121 MB RAM, roughly 317 Hz theoretical throughput, and detection maintained up to 15 m.
- The released dataset, open-source labeling tool, and trained models constitute a practical reproducible baseline for resource-constrained FSD teams.
- Traditional machine learning remains competitive with deep learning for resource-constrained real-time perception.
Strengths
Leakage-aware evaluation protocol
Section 3.1 keeps clusters from the same run entirely within one split and retains only every 10th LiDAR frame, directly addressing the temporal-correlation leakage that undermines many small perception datasets.
Quantified CPU runtime with a real ablation
Table 3 shows end-to-end runtime falling 4.54 → 3.50 → 3.13 ms as features drop 9 → 8 → 7 while F1 rises 97.77% → 98.33%; the unmeasured 12-feature runtime is honestly marked (†) rather than interpolated.
Strong reproducibility package
Honest limitation disclosure
Thoughtful viewpoint-invariance design
Weaknesses
No same-data baselines
Statistical thinness
Internal runtime inconsistency
Prevalence shift without stratified reporting
Single-source dataset
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets Formula Student Driverless (FSD) teams without GPU budgets, proposing a LiDAR-only, CPU-only cone perception pipeline: ray-based ground removal (Eqs. 1–4), IMU-based motion compensation (Eqs. 5–6), DBSCAN clustering, and a Random Forest over seven geometric features (Table 1). The four stated contributions are importance-based feature reduction (12 → 7 features), an open-source labeling/training tool (Cone-Cluster-Labeler [16]), a manually annotated dataset of 2,371 clusters from real FSD events, and reproducible trained models. Central quantitative claims: 98.33% F1 (99.16% precision, 97.51% recall) on a 504-sample held-out test split; 3.13 ms end-to-end CPU runtime (~317 Hz theoretical vs a 20 Hz sensor); ~121 MB RAM; detection up to 15 m. Claims are mostly scoped honestly, with limitations disclosed in Section 5.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is this LiDAR cone detection method?
On a held-out test set of 504 clusters collected at real Formula Student Driverless events, the pipeline reports 99.16% precision, 97.51% recall, and a 98.33% F1-score for separating cones from non-cones, with detection maintained up to a 15 m range using an Ouster OS1-32 sensor.
Does this cone detector need a GPU?
No. The entire pipeline — ground removal, IMU motion compensation, DBSCAN clustering, and Random Forest classification — runs on CPU only. The authors measured 3.13 ms end-to-end on an Intel Core Ultra 9 285H (roughly 317 Hz theoretical throughput) using about 121 MB of RAM.
Is the dataset and code available?
Yes, according to the paper: a dataset of 2,371 manually labeled LiDAR clusters from real FSD events, the open-source Cone-Cluster-Labeler annotation and training tool on GitHub, and the trained Random Forest models are all released for other teams to use.
What are the main limitations?
Detection degrades beyond 15 m because the 32-beam sensor returns fewer than five points per cone; rain and fog were not tested; the dataset comes from a single sensor and team; and the paper reports no statistical uncertainty and no head-to-head comparison with deep learning on the same data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Mathematical symbols in Section 3.2 (Eqs. 1–6) and the Table 1 formulas render as placeholder glyphs in the extracted text, so exact equation forms could not be verified; structure was inferred from surrounding prose.
- Fig. 2 confusion-matrix cell values are not present in the extracted text (figure content is not extracted), so class-wise error counts could not be independently checked against Table 3.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.