Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BikeScenes: Online LiDAR Semantic Segmentation for Bicycles

Denniz Goren, Holger Caesar

TL;DR

BikeScenes tackles the challenge of enabling 3D LiDAR semantic segmentation on a bicycle by introducing BikeScenes-lidarseg, a bicycle-perspective dataset with 29 classes, and demonstrating the necessity of domain adaptation through fine-tuning on-bike. The authors select an efficient on-board architecture (FRNet), establish a robust training regimen, and show a substantial mIoU improvement from 13.8% (SemanticKITTI pre-training only) to 63.6% after fine-tuning on BikeScenes. They map BikeScenes to SemanticKITTI for comparability, analyze domain gaps, and integrate a ROS 2 inference pipeline on the SenseBike, including AMP-based speed-ups. The work provides a practical resource and validated workflow toward cyclist-centric perception, with implications for on-bike ADAS-like features and safer urban cycling.

Abstract

The vulnerability of cyclists, exacerbated by the rising popularity of faster e-bikes, motivates adapting automotive perception technologies for bicycle safety. We use our multi-sensor 'SenseBike' research platform to develop and evaluate a 3D LiDAR segmentation approach tailored to bicycles. To bridge the automotive-to-bicycle domain gap, we introduce the novel BikeScenes-lidarseg Dataset, comprising 3021 consecutive LiDAR scans around the university campus of the TU Delft, semantically annotated for 29 dynamic and static classes. By evaluating model performance, we demonstrate that fine-tuning on our BikeScenes dataset achieves a mean Intersection-over-Union (mIoU) of 63.6%, significantly outperforming the 13.8% obtained with SemanticKITTI pre-training alone. This result underscores the necessity and effectiveness of domain-specific training. We highlight key challenges specific to bicycle-mounted, hardware-constrained perception systems and contribute the BikeScenes dataset as a resource for advancing research in cyclist-centric LiDAR segmentation.

BikeScenes: Online LiDAR Semantic Segmentation for Bicycles

TL;DR

BikeScenes tackles the challenge of enabling 3D LiDAR semantic segmentation on a bicycle by introducing BikeScenes-lidarseg, a bicycle-perspective dataset with 29 classes, and demonstrating the necessity of domain adaptation through fine-tuning on-bike. The authors select an efficient on-board architecture (FRNet), establish a robust training regimen, and show a substantial mIoU improvement from 13.8% (SemanticKITTI pre-training only) to 63.6% after fine-tuning on BikeScenes. They map BikeScenes to SemanticKITTI for comparability, analyze domain gaps, and integrate a ROS 2 inference pipeline on the SenseBike, including AMP-based speed-ups. The work provides a practical resource and validated workflow toward cyclist-centric perception, with implications for on-bike ADAS-like features and safer urban cycling.

Abstract

The vulnerability of cyclists, exacerbated by the rising popularity of faster e-bikes, motivates adapting automotive perception technologies for bicycle safety. We use our multi-sensor 'SenseBike' research platform to develop and evaluate a 3D LiDAR segmentation approach tailored to bicycles. To bridge the automotive-to-bicycle domain gap, we introduce the novel BikeScenes-lidarseg Dataset, comprising 3021 consecutive LiDAR scans around the university campus of the TU Delft, semantically annotated for 29 dynamic and static classes. By evaluating model performance, we demonstrate that fine-tuning on our BikeScenes dataset achieves a mean Intersection-over-Union (mIoU) of 63.6%, significantly outperforming the 13.8% obtained with SemanticKITTI pre-training alone. This result underscores the necessity and effectiveness of domain-specific training. We highlight key challenges specific to bicycle-mounted, hardware-constrained perception systems and contribute the BikeScenes dataset as a resource for advancing research in cyclist-centric LiDAR segmentation.

Paper Structure

This paper contains 20 sections, 2 equations, 12 figures, 3 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Outdoor LiDAR Segmentation Datasets
  4. LiDAR Segmentation Architectures
  5. BIKESCENES LIDARSEG DATASET
  6. Recording
  7. Labeling Procedure
  8. Subsequences - Train/Val/Test Split
  9. EXPERIMENTS AND RESULTS
  10. Evaluation Metrics
  11. Architecture Selection for On-bike Inference
  12. Class Remapping
  13. Domain Gap Analysis and Training Schedules
  14. Performance on the Test Set
  15. Confusion Matrix Analysis
  16. ...and 5 more sections

Figures (12)

  • Figure 1: Scene from the BikeScenes-lidarseg Dataset.
  • Figure 2: The SenseBike and sensors used to capture BikeScenes.
  • Figure 3: Map of aggregated labeled scans. building; road; sidewalk; vegetation; bike-path.
  • Figure 4: GPS trajectory and subsequence categorization.
  • Figure 5: Class distribution of the BikeScenes dataset. The number of points for dynamic classes is divided between non-moving (solid bars) and moving objects (hatched bars).
  • ...and 7 more figures