Robust Object Segmentation Using a Multi-Layer Laser Scanner

Beomseong Kim, Baehoon Choi, Minkyun Yoo, Hyunju Kim, Euntai KimSensors 2014 · DOI

TL;DR

This paper introduces a novel robust object segmentation method for multi-layer laser scanners, crucial for advanced driver assistance systems (ADAS). It effectively addresses the challenge of "ghost" detections caused by reflections, significantly improving the reliability of object recognition in real-world driving scenarios.

What Problem Does This Paper Address?

Advanced Driver Assistance Systems (ADAS) rely heavily on accurately perceiving the surrounding environment. While various sensors like radar and cameras have limitations (e.g., inaccurate distance for radar, unreliable distance for cameras), laser scanners offer a compelling alternative due to their accurate range and bearing measurements. However, a significant challenge in using laser scanners, particularly multi-layer ones, is the generation of "ghost" segments. These are false detections caused by reflections from surfaces like the ground, or from lights in fog or rain. These ghosts can mislead the ADAS, potentially leading to dangerous situations. Existing segmentation methods, often designed for single-layer scanners, struggle to effectively filter these ghosts, degrading the overall system's reliability.

Methods

The paper proposes a two-stage approach:

  1. Adaptive Breakpoint Detector (ABD) Segmentation for Multi-Layer Laser Scanners: This method extends the traditional breakpoint detection (BD) approach to handle the complexities of multi-layer laser scanners. Key innovations include:
  2. Considering Layer Information: Unlike single-layer methods, this approach accounts for the layer information of each point, recognizing that points at the same bearing but different layers can belong to distinct objects (e.g., a sedan closer than a bus).
  3. Handling Incomplete Data: It addresses the inherent sparsity and non-uniformity of data in the angular-layer plane (θ-l plane) typical of multi-layer scanners.
  4. O(N) Complexity: The algorithm is designed for efficient, real-time processing with a linear time complexity with respect to the number of measurement points (N). It achieves this by maintaining a candidate set of the newest points in each layer and testing connectivity only against this set.
  5. Robust Segmentation through Ghost Elimination: This is the core contribution for improving reliability. It builds upon the ABD method with two key modifications:
  6. Limited Range Processing: For points within a limited distance (40m), connectivity with points on the same layer is explicitly skipped. This is based on the observation that ghosts often appear on a single layer and are detected at close range. By preventing same-layer connections for close points, ghost segments are less likely to form meaningful objects.
  7. Expanded Candidate Set: The candidate set (M) for connectivity testing is expanded to include up to two of the newest points from each layer (totaling up to 8 points), allowing for more robust object detection even with sparse data.
  8. Small Segment Elimination: Finally, small segments (those with fewer than a minimum number of points) are eliminated, further filtering out spurious detections.

Datasets

The paper does not explicitly mention training datasets. The evaluation is performed using data collected from a Kia K900 vehicle equipped with an IBEO LUX2010 multi-layer laser scanner and a camera. The experiments cover various real-world scenarios, including uphill roads, flat roads, rainy weather, and foggy weather, to test the robustness of the proposed method.

Key Insights and Ideas

  • Ghosts are a critical problem: The paper clearly demonstrates that ghost segments, especially those from ground reflections or fog, pose a significant risk to ADAS safety.
  • Multi-layer scanners have unique challenges: The inherent structure of multi-layer data (points at the same bearing but different layers) requires specialized segmentation beyond single-layer approaches.
  • Ghosts are often close-range and single-layer: This observation is the foundation for the robust ghost elimination strategy, leading to a significant improvement in segmentation quality.
  • Balancing ghost elimination and inlier survival is key: The proposed method aims to remove ghosts without discarding legitimate object measurements, a crucial trade-off for ADAS.

Key Results

The proposed robust segmentation method significantly outperforms the standard ABD segmentation in ghost elimination across various conditions.

  • Ghost Elimination Ratio: The robust method achieved over 95% ghost elimination in a robust manner, regardless of weather or road conditions. In contrast, the ABD method's ghost elimination ratio varied widely from 17% to 65%.
  • Performance in Challenging Conditions: The robust method effectively removed ghosts in uphill roads and rainy conditions where the ABD method struggled.
  • Inlier Survival Ratio: Both methods maintained sufficiently high inlier survival ratios, indicating that important object measurements are not accidentally removed.
  • Computation Time: The robust method incurred a slightly longer computation time compared to ABD, but the difference was deemed not significant for real-time applications.

Limitations and Future Work

  • Shallow Obstacles: The paper acknowledges that if an obstacle is very shallow and detected on only a single layer, it might not be detected by the robust method. However, they argue that the scanner's resolution makes this rare for obstacles within a reasonable distance.
  • Limited Range for Robustness: The robust ghost elimination strategy is applied within a limited 40m range. For distances beyond 40m, the standard ABD segmentation is used. This suggests a potential area for further research to extend the robust ghost elimination to longer ranges.
  • Further Quantitative Analysis: While tables provide quantitative results, the paper could benefit from more detailed statistical analysis or a larger dataset for broader validation.

Summary updated

BibTeX
@article{kim2014robust,
  author = {Beomseong Kim and Baehoon Choi and Minkyun Yoo and Hyunju Kim and Euntai Kim},
  title = {{Robust Object Segmentation Using a Multi-Layer Laser Scanner}},
  journal = {Sensors},
  year = {2014},
  doi = {10.3390/s141120400},
}
Figure