Unsupervised Feature Learning for 3D Scene Reconstruction with Occupancy Maps

Vitor Campanholo Guizilini, Fabio Tozeto Ramos

Abstract

This paper presents an unsupervised feature learning pipeline designed to compress massive 3D point clouds into high-level geometric features for efficient robotic mapping. By introducing a novel clustering algorithm, ASK-means, and an orientation-based merging strategy, the authors create 3D occupancy maps that are orders of magnitude smaller in memory and significantly faster to process than state-of-the-art methods. The resulting Planar Surface Hilbert Maps (PS-HM) maintain or improve reconstruction accuracy while enabling real-time performance on mobile robotic platforms.

Introduction

In mobile robotics, sensors like LiDAR and RGB-D cameras generate millions of data points every second. While this provides incredible detail, it creates a massive computational burden for storage, processing, and real-time navigation. Traditional occupancy mapping techniques often treat these points individually or within fixed-size voxels, ignoring the structured patterns—such as flat walls, floors, and ceilings—prevalent in human-made environments. This research addresses the challenge of automatically learning to group raw sensor data into semantically meaningful planar features without human supervision.

Methods

The authors propose a two-step unsupervised feature learning pipeline integrated into the Hilbert Maps framework:

  1. Spatial Clustering (ASK-means): The "Automatic Stop K-means" algorithm automatically determines the optimal number of clusters ($k$). It utilizes a novel initialization strategy that removes points from the candidate pool once they are near a newly selected cluster center, accelerating the process by approximately 40% compared to the standard K-means++ algorithm.
  2. Orientation Clustering: Following spatial grouping, the algorithm analyzes the statistical covariance of each group. Neighboring clusters with similar normal vectors are merged into single "super-clusters" representing continuous surfaces.
  3. 3D Alignment and Modeling: To ensure precise orientation, the system uses Rodrigues’ rotation formula to optimize cluster axes. A specialized planar covariance function is then introduced as a kernel for Hilbert Maps, allowing the system to represent large surfaces with minimal parameters.
Performance comparison of ASK-means versus k-means++ in terms of potential and computation time.

Experimental Setup

The approach was evaluated across three distinct environments:

  • Room: A simulated indoor environment for controlled testing.
  • Corridor: A real-world indoor dataset capturing complex architectural features.
  • Outdoor: A real-world dataset containing both structured buildings and unstructured objects like trees and bushes.

Key Results and Discussion

The primary breakthrough is the scale of data reduction achieved. In the "Room" dataset, the algorithm reduced 582,463 raw points down to just 42 planar clusters—a reduction of four orders of magnitude. Despite this compression, the proposed PS-HM achieved higher classification accuracy than industry standards like Octo-Map, particularly when training data was sparse.

Extracted planar features from a raw dataset, demonstrating the algorithm's ability to identify and group structured surfaces.

Key insights include:

  • Shape Learning: By learning the covariance (shape) of features rather than just their position, the map can represent large wall sections with a single parameter.
  • Hybrid Robustness: The system identifies planes where they exist but gracefully reverts to standard modeling for non-planar objects, ensuring reliability in diverse environments.
  • Efficiency: While orientation clustering adds a small initial overhead, it results in a much smaller feature vector, making subsequent mapping and querying phases significantly faster for real-time robotics.

Conclusion

This work demonstrates that unsupervised feature learning can transform massive point clouds into efficient, high-level geometric maps. While currently optimized for planar surfaces, future work will focus on detecting and modeling other feature types, such as cylinders and spheres, using similar statistical methods. Additionally, the authors aim to further optimize clustering speed for deployment on low-power embedded systems found in small-scale mobile robots.

Summary updated

Abstract

This paper addresses the task of unsupervised feature learning for three-dimensional occupancy mapping, as a way to segment higher-level structures based on raw unorganized point cloud data. In particular, we focus on detecting planar surfaces, which are common in most structured or semi-structured environments. This segmentation is then used to minimize the amount of parameters necessary to properly create a 3D occupancy model of the surveyed space, thus increasing computational speed and decreasing memory requirements. As the 3D modeling tool, an extension to Hilbert Maps was selected, since it naturally uses a feature-based representation of the environment to achieve real-time performance. Experiments conducted in simulated and real large-scale datasets show a substantial gain in performance, while decreasing the amount of stored information by orders of magnitude without sacrificing accuracy.
BibTeX
@inproceedings{guizilini2017unsupervised,
  author = {Vitor Campanholo Guizilini and Fabio Tozeto Ramos},
  title = {{Unsupervised Feature Learning for 3D Scene Reconstruction with Occupancy Maps}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.1609/aaai.v31i1.11039},
}
Figure