MARF: Multiscale Adaptive-Switch Random Forest for Leg Detection With 2-D Laser Scanners

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Leg detection is a critical initial step for 2-D laser-based people detection and tracking.
• Existing leg detectors struggle with noise and multiscale variations inherent in laser scan point clouds, leading to performance degradation.

Purpose of the Study:

• To propose a novel Multiscale Adaptive-Switch Random Forest (MARF) for robust leg detection.
• To address the challenges of noise and multiscale characteristics in 2-D laser scans for improved people detection and tracking.

Main Methods:

• Developed an adaptive-switch decision tree combining noise-sensitive weighted classification and noise-invariant binary classification for enhanced robustness.
• Designed a multiscale random forest structure to effectively detect legs at varying distances, accounting for point cloud sparsity.
• Implemented the MARF for leg detection, enabling sparser human leg discovery from point clouds.

Main Results:

• The MARF detector demonstrated superior robustness to noise compared to existing methods.
• Achieved improved performance on the Moving Legs dataset, outperforming state-of-the-art leg detectors.
• The MARF pipeline maintained a speed of over 60 FPS on low-computational laptops.
• Significant performance gains were observed when applying MARF to a complete people detection and tracking system.

Conclusions:

• The proposed MARF significantly enhances leg detection accuracy and robustness in challenging 2-D laser scan environments.
• MARF offers a computationally efficient and effective solution for real-time people detection and tracking systems.
• This method provides a foundation for more reliable robotic perception and surveillance applications.