Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LiDAR Based Semantic Perception for Forklifts in Outdoor Environments

Benjamin Serfling, Hannes Reichert, Lorenzo Bayerlein, Konrad Doll, Kati Radkhah-Lens

TL;DR

This paper tackles semantic perception for autonomous forklifts operating in outdoor industrial environments by proposing a dual-LiDAR system that provides broad spatial coverage and short-range precision. The authors develop a lightweight CNN-based segmentation pipeline that uses a spherical LiDAR representation, surface normals, and reflectivity, fused across two synchronized LiDARs, and enhanced by an attention-based neck and feature pyramid network to achieve real-time performance. They train with a combined cross-entropy and Tversky loss, pretrain on a large automotive dataset, and finetune on a custom industrial dataset featuring forklifts, pedestrians, and ground/lanes. The evaluation shows that EfficientNet-based backbones offer a favorable accuracy–latency trade-off under a $33.3$ ms budget, achieving up to $\text{mIoU}=74.14\%$ while maintaining real-time operation, and demonstrates strong segmentation of drivable ground, lane markings, and safety-critical objects. The work provides a practical path toward safety-aware, fully autonomous forklift navigation in dynamic warehouse and outdoor yard environments, with future directions including temporal fusion and multi-modal sensing.

Abstract

In this study, we present a novel LiDAR-based semantic segmentation framework tailored for autonomous forklifts operating in complex outdoor environments. Central to our approach is the integration of a dual LiDAR system, which combines forward-facing and downward-angled LiDAR sensors to enable comprehensive scene understanding, specifically tailored for industrial material handling tasks. The dual configuration improves the detection and segmentation of dynamic and static obstacles with high spatial precision. Using high-resolution 3D point clouds captured from two sensors, our method employs a lightweight yet robust approach that segments the point clouds into safety-critical instance classes such as pedestrians, vehicles, and forklifts, as well as environmental classes such as driveable ground, lanes, and buildings. Experimental validation demonstrates that our approach achieves high segmentation accuracy while satisfying strict runtime requirements, establishing its viability for safety-aware, fully autonomous forklift navigation in dynamic warehouse and yard environments.

LiDAR Based Semantic Perception for Forklifts in Outdoor Environments

TL;DR

This paper tackles semantic perception for autonomous forklifts operating in outdoor industrial environments by proposing a dual-LiDAR system that provides broad spatial coverage and short-range precision. The authors develop a lightweight CNN-based segmentation pipeline that uses a spherical LiDAR representation, surface normals, and reflectivity, fused across two synchronized LiDARs, and enhanced by an attention-based neck and feature pyramid network to achieve real-time performance. They train with a combined cross-entropy and Tversky loss, pretrain on a large automotive dataset, and finetune on a custom industrial dataset featuring forklifts, pedestrians, and ground/lanes. The evaluation shows that EfficientNet-based backbones offer a favorable accuracy–latency trade-off under a ms budget, achieving up to while maintaining real-time operation, and demonstrates strong segmentation of drivable ground, lane markings, and safety-critical objects. The work provides a practical path toward safety-aware, fully autonomous forklift navigation in dynamic warehouse and outdoor yard environments, with future directions including temporal fusion and multi-modal sensing.

Abstract

In this study, we present a novel LiDAR-based semantic segmentation framework tailored for autonomous forklifts operating in complex outdoor environments. Central to our approach is the integration of a dual LiDAR system, which combines forward-facing and downward-angled LiDAR sensors to enable comprehensive scene understanding, specifically tailored for industrial material handling tasks. The dual configuration improves the detection and segmentation of dynamic and static obstacles with high spatial precision. Using high-resolution 3D point clouds captured from two sensors, our method employs a lightweight yet robust approach that segments the point clouds into safety-critical instance classes such as pedestrians, vehicles, and forklifts, as well as environmental classes such as driveable ground, lanes, and buildings. Experimental validation demonstrates that our approach achieves high segmentation accuracy while satisfying strict runtime requirements, establishing its viability for safety-aware, fully autonomous forklift navigation in dynamic warehouse and yard environments.

Paper Structure

This paper contains 21 sections, 4 equations, 8 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Motivation
  3. Requirements
  4. Real-Time Constraints
  5. Segmentation Accuracy
  6. Related Work
  7. Automotive LiDAR Semantic Segmentation
  8. LiDAR-Based Perception in Material Handling
  9. Contributions
  10. Outline
  11. Dataset
  12. Method
  13. Preprocessing
  14. Semantic Segmentation
  15. Implementation Details
  16. ...and 6 more sections

Figures (8)

  • Figure 1: Prototype forklift in action. © Linde Material Handling
  • Figure 2: Distributions of semantic classes in SemanticTHAB.
  • Figure 3: Reference Map of Scene 0000.
  • Figure 4: The forklift vehicle frame (following ISO 8855 conventions) and the LiDAR sensor coordinate frame.
  • Figure 5: Spherical projections of reflectivity measurements (top), semantic annotations (middle), and surface normals (bottom)
  • ...and 3 more figures