Clustering and visualization of non-classified points from LiDAR data for helicopter navigation

Ferdinand Eisenkeil, Tobias Schafhitzel, Uwe Kühne, Oliver DeußenProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2014 · DOI

TL;DR

This paper presents a novel method for clustering and visualizing unclassified LiDAR data to enhance helicopter navigation safety. By dynamically fusing clusters from sequential sensor frames, the system provides a stable, non-distracting representation of potential obstacles, crucial for pilots during challenging flight conditions.

What Problem Does This Paper Address?

Helicopter pilots face significant challenges in low-visibility conditions, such as dust, snow, or night flights, where unclassified obstacles like wires can pose a severe crash risk. Existing synthetic vision systems can classify common objects like ground or trees, but they often struggle to effectively represent and visualize "unclassified" points, which represent potential hazards. The challenge is to process dense, rapidly acquired LiDAR data to identify and present these unclassified dangers to the pilot in a clear, intuitive, and non-overwhelming manner, without cluttering the display.

Methods

The core of the proposed method involves a dynamic, density-based clustering approach using OPTICS, chosen for its ability to handle varying data densities common in LiDAR scans. Here's a breakdown of their key techniques:

  • Dynamic Clustering with OPTICS: Instead of processing each LiDAR frame in isolation, the system uses OPTICS to cluster unclassified points. OPTICS is favored because LiDAR data density decreases with distance from the helicopter, and OPTICS can adapt to these density variations.
  • Convex Hull Representation: For each identified cluster, a convex hull is calculated using the Graham Scan algorithm. This provides a simplified geometric representation of the cluster's extent.
  • Cluster Fusion using R-functions: This is a key innovation. To achieve temporal stability and a smooth visualization, clusters from consecutive sensor frames are fused. This is accomplished by converting convex hulls into implicit functions and then applying R-functions, which enable Boolean operations (like union and intersection) on these implicit functions. This allows for flexible merging of clusters based on their spatial relationships and context.
  • Visualization: The fused convex hulls are then visualized. The goal is to provide a simple, highlighted representation (e.g., lines) of threatening unclassified points, rather than displaying all raw LiDAR points, which would be overwhelming. The visualization is designed to be adaptable for both head-down and head-mounted displays.

Datasets

The paper primarily discusses the use of LiDAR sensor data recorded in short time intervals. They also mention using synthetic data with characteristics similar to real LiDAR data (stable helicopter position, no noise bias, same spatial/temporal resolution) for evaluation and comparison. No specific named datasets are mentioned.

Key Insights and Ideas

  • Temporal Stability is Crucial: Simply clustering each LiDAR frame independently leads to flickering and unstable visualizations. The paper emphasizes that fusing clusters over time is essential for a pilot-friendly display.
  • R-functions for Flexible Fusion: Using R-functions for Boolean operations on implicit representations of convex hulls offers a more sophisticated and flexible way to merge clusters compared to simple binary operations. This allows for context-aware fusion, adapting to whether a cluster is fully within the sensor's view or at its border.
  • Abstraction for Clarity: The decision to represent clusters by their convex hulls and then visualize these hulls (rather than raw points) is a critical insight for reducing visual clutter and improving pilot comprehension.
  • Real-time Capability: The system is designed to operate within the 300ms time frame between LiDAR frames, making it suitable for real-time helicopter navigation.

Key Results

  • The proposed approach generates a stable visualization of unclassified LiDAR data over time, unlike simpler methods like basic OPTICS clustering applied frame-by-frame.
  • The system is demonstrated to be real-time capable, processing data within the required 300ms interval.
  • Comparisons with a basic OPTICS approach using synthetic data show that the proposed method results in smoother contours and more stable clustering with minimal changes between frames.
  • The visualization method is adaptable for both head-down and head-mounted displays, offering potential for improved situational awareness.

Limitations and Future Work

  • Initial DBSCAN Instability: The authors mention that an initial implementation using DBSCAN was too slow and unstable over time, leading them to switch to OPTICS. While OPTICS improved speed, it still presented stability challenges that their fusion method aims to address.
  • Visualization Evaluation: The paper notes that the specific representation of unclassified data needs further evaluation. They suggest a potential application case of visualizing only unclassified pixels on the landing site to avoid unnecessary clutter.
  • Future Work: The authors intend to further evaluate the proposed visualization methods and explore more refined ways to represent unclassified data, potentially focusing on specific areas like the landing zone. They also aim to integrate this system into a broader synthetic vision processing chain.

Summary updated

Abstract

In this paper we propose a dynamic DBSCAN-based method to cluster and visualize unclassified and potential dangerous obstacles in data sets recorded by a LiDAR sensor. The sensor delivers data sets in a short time interval, so a spatial superposition of multiple data sets is created. We use this superposition to create clusters incrementally. Knowledge about the position and size of each cluster is used to fuse clusters and the stabilization of clusters within multiple time frames. Cluster stability is a key feature to provide a smooth and un-distracting visualization for the pilot. Only a few lines are indicating the position of threatening unclassified points, where a hazardous situation for the helicopter could happen, if it comes too close. Clustering and visualization form a part of an entire synthetic vision processing chain, in which the LiDAR points support the generation of a real-time synthetic view of the environment.
BibTeX
@inproceedings{eisenkeil2014clustering,
  author = {Ferdinand Eisenkeil and Tobias Schafhitzel and Uwe Kühne and Oliver Deußen},
  title = {{Clustering and visualization of non-classified points from LiDAR data for helicopter navigation}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2014},
  doi = {10.1117/12.2050497},
}
Figure