Vehicle Detection Based on Probability Hypothesis Density Filter

Feihu Zhang, Alois C. Knoll

TL;DR

This paper introduces a novel LiDAR-based vehicle detection system that leverages the Probability Hypothesis Density (PHD) filter to overcome data association challenges in complex urban environments. By combining a Random Hypersurface Model (RHM) for object hypothesis generation with a Support Vector Machine (SVM) for verification, the system achieves robust vehicle detection even with occlusions and varying lighting conditions.

What Problem Does This Paper Address?

Traffic accidents remain a significant global health concern, and autonomous driving systems are seen as a key solution to mitigate them. A crucial component of these systems is accurate and reliable vehicle detection. While vision-based methods are common, they struggle with varying light conditions and have limited fields of view, leading to false positives. LiDAR sensors offer robustness to lighting and provide depth information, but they generate a large number of measurements without inherent association information, making it difficult to track objects and distinguish between them. This paper tackles the challenge of robust vehicle detection using LiDAR data, specifically addressing the problem of unknown data association in complex environments.

Methods

The proposed approach consists of two main phases: hypothesis generation and hypothesis verification.

  • Hypothesis Generation (RHM-GM-PHD Filter):
    • Random Hypersurface Model (RHM): Objects are represented by their shape parameters using Fourier series expansion. This allows for a compact representation of object contours.
  • Gaussian Mixture Probability Hypothesis Density (GM-PHD) Filter: This is the core of the hypothesis generation. The PHD filter is a probabilistic approach that handles multiple targets and unknown data association by propagating a "density" of potential targets. The paper adapts a standard GM-PHD filter for "Extended Targets" (ETs) to handle the multiple measurements originating from a single object's surface, which is typical for LiDAR data. This ET-GM-PHD filter directly processes LiDAR measurements without requiring prior clustering or association steps.
  • Hypothesis Verification (SVM):
    • Support Vector Machine (SVM): After the PHD filter generates potential object hypotheses (represented by their shape parameters, i.e., Fourier coefficients), an SVM is used to classify these hypotheses as either vehicles or non-vehicles. This helps to eliminate false positives generated during the hypothesis generation phase. The SVM is trained using Fourier coefficients extracted from labeled objects.

Datasets

The paper utilizes the KITTI dataset for both training and evaluation. Specifically, it uses 5000 training frames with 1893 manually labeled objects for training the SVM classifier. An additional 2000 frames are used for testing.

Key Insights and Ideas

  • LiDAR for Robustness: The paper highlights the advantage of LiDAR in overcoming the limitations of vision-based systems, particularly in challenging lighting conditions.
  • PHD Filter for Data Association: The core innovation is the application of the PHD filter, specifically an Extended Target (ET) variant, to directly handle the high volume of LiDAR measurements without explicit data association. This is crucial for real-time applications.
  • Shape as a Discriminative Feature: Using Fourier coefficients derived from the RHM as shape descriptors for classification is a novel contribution. This allows the system to distinguish between different object types based on their geometric form.
  • Two-Phase Approach: The combination of a probabilistic hypothesis generation phase (PHD filter) and a discriminative verification phase (SVM) provides a robust pipeline for vehicle detection.

Key Results

  • Performance: The proposed approach achieves high performance in "easy" scenarios (fully visible vehicles with uniform reflections), outperforming state-of-the-art methods by a significant margin (almost 20%-40%).
  • Speed: The average processing time is 5 seconds per frame, implemented in Matlab on a dual-core 3 GHz processor.
  • Limitations in Occlusion: The performance drops significantly in "moderate" and "hard" scenarios (partially occluded or difficult-to-see vehicles). This is attributed to the influence of invisible measurements on the Fourier coefficients used by the SVM.

Limitations and Future Work

  • Occlusion Handling: The paper explicitly acknowledges that the performance degrades in scenarios with significant occlusion. The SVM classification is particularly affected as it relies on visible measurements to compute Fourier coefficients.
  • Future Work: The authors suggest that future work should focus on improving the detection of occluded vehicles. This could involve developing more robust shape descriptors or enhancing the PHD filter's ability to handle partial observations.

Summary updated

BibTeX
@inproceedings{zhang2016vehicle,
  author = {Feihu Zhang and Alois C. Knoll},
  title = {{Vehicle Detection Based on Probability Hypothesis Density Filter}},
  booktitle = {Sensors},
  year = {2016},
  url = {https://openreview.net/forum?id=VH_U7x2Fh_W},
}
Figure