Height Change Feature Based Free Space Detection

Steven Schreck, Hannes Reichert, Manuel Hetzel, Konrad Doll, Bernhard Sick

Abstract

Autonomous Guided Vehicles (AGVs), particularly forklifts, require high-speed and reliable free space detection to navigate unstructured industrial environments safely. This paper presents a novel heuristic method that treats 3D LiDAR point clouds as 2D spherical projections, applying convolutional gradient filters to estimate surface normals and segment the ground. By bypassing the need for labeled training data and heavy computational models, the proposed approach achieves processing speeds up to 105 Hz, significantly outperforming existing machine learning and heuristic baselines while maintaining competitive accuracy.

Introduction

In the demanding environments of factory floors and production halls, autonomous forklifts must navigate areas filled with moving workers, industrial trucks, and obstacles like pallets and racks. Unlike autonomous cars on structured roads, these vehicles often operate without the benefit of lane markings or predictable traffic patterns.

Existing solutions for free space detection—identifying drivable areas—typically fall into two categories: computationally intensive Machine Learning (ML) methods that require massive labeled datasets, or faster heuristic methods that often struggle with sensor noise or uneven terrain. This research bridges the gap by introducing a high-speed, geometry-based heuristic designed for real-time collision avoidance in busy industrial settings.

Methodology

The core innovation of this approach is the simplification of 3D geometry into a 2D image processing problem. By restructuring the LiDAR data, the authors leverage hardware-friendly convolutions to achieve high-frequency updates. The pipeline consists of four primary stages:

1. Spherical Projection

The raw 3D LiDAR point cloud is first converted into a "staggered" 2D image. In this representation, each pixel corresponds to a specific laser beam's measurement. This ensures that neighboring pixels in the image correspond to neighboring points in the physical world, allowing for efficient local spatial analysis.

2. Fast Surface Normal Estimation

Rather than employing traditional, computationally expensive mathematical approaches like least-squares plane fitting, the method applies Scharr filters. These convolutional gradient filters, typically used for edge detection in computer vision, are used here to calculate surface normals (the direction a surface faces) at extremely high speeds.

3. Height Change Features

Once the surface normals are estimated, the points are transformed into the vehicle’s coordinate system. The algorithm calculates the angle between the surface normal and the vertical z-axis. For a surface to be considered flat and drivable, its normal must point substantially upward, aligning with the vertical axis.

4. Statistical Filtering

To distinguish the actual floor from other flat surfaces—such as the tops of crates or pallets—the system applies a statistical $1\sigma$ filter based on height distribution. This step isolates the dominant ground plane and filters out elevated flat surfaces that are not drivable.

Experimental Evaluation

The authors evaluated the algorithm using two distinct datasets to test its versatility and robustness:

  • Semantic KITTI: A standard outdoor benchmark for autonomous driving. To establish a ground truth for free space, the authors aggregated labels for roads, sidewalks, and parking areas.
  • Factory Site Dataset: A custom dataset collected using a forklift equipped with two 128-layer Ouster OS0 sensors. This dataset includes indoor factory floors and outdoor asphalt roads with longitudinal slopes of up to 12%.

Results and Discussion

Speed and Efficiency

The method achieved a processing speed of 105 Hz on an Intel i7 CPU when tested on the KITTI dataset. This performance is roughly twice as fast as the leading heuristic baseline (LineFit) and nearly seven times faster than state-of-the-art ML models like 2DPASS, which operates at approximately 16 Hz. This low latency is critical for "pick and place" operations where the vehicle must react instantly to dynamic obstacles.

Accuracy

  • Semantic KITTI: The method reached a mean Intersection over Union (mIoU) of 50.90%, outperforming other heuristic methods which ranged between 32% and 43%.
  • Factory Site: The algorithm performed even better in the industrial environment, achieving an mIoU of 63.30% at 54 Hz (per LiDAR). The higher vertical resolution of the 128-layer Ouster sensors compared to the 64-layer Velodyne used in KITTI contributed to cleaner surface normals and higher precision.

Key Insights

The research demonstrates that for specific industrial tasks, well-tuned geometric heuristics can outperform complex neural networks in both speed and deployment flexibility. Because the system does not require a training phase, it can be deployed across different vehicle types or new environments without the need for costly and time-consuming data labeling.

Limitations and Future Work

While highly effective in structured and semi-structured environments, the method is sensitive to high-noise areas. For instance, dense vegetation in the KITTI dataset produced erratic surface normals, which reduced accuracy compared to the cleaner factory settings.

Future research will focus on improving ramp and slope handling to better distinguish between steep inclines and obstacles. Additionally, the authors plan to integrate the "non-ground" points identified by this system directly into object detection and tracking pipelines. There is also significant potential for further speed gains by implementing the convolutional filters on a GPU.

Conclusion

This height change feature-based detection method provides a robust, real-time solution for autonomous forklift navigation. By treating 3D spatial data as a 2D filtering problem, the approach delivers the high-frequency updates necessary for safe industrial operation without the computational overhead of deep learning.

Summary updated

BibTeX
@inproceedings{schreck2023height,
  author = {Steven Schreck and Hannes Reichert and Manuel Hetzel and Konrad Doll and Bernhard Sick},
  title = {{Height Change Feature Based Free Space Detection}},
  booktitle = {ICCMA},
  year = {2023},
  doi = {10.1109/iccma59762.2023.10374705},
}
Figure