On-Board Object Detection: Multicue, Multimodal, and Multiview Random Forest of Local Experts.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Object detection in real-world scenarios remains challenging despite advances.
• Effective detectors require multiple cues, imaging modalities, and robust multiview (MV) classification.
• Integrating these components is crucial for improving accuracy.

Purpose of the Study:

• To evaluate the individual and combined impact of multicue, multimodality, and MV classification on object detection accuracy.
• To explore the fusion of RGB and depth maps from high-definition light detection and ranging (LiDAR) for enhanced detection.
• To identify the most effective strategies for improving object detection performance in challenging conditions.

Main Methods:

• Extensive evaluation of object detection components: multicue, multimodality, and MV classifier.
• Exploration of RGB and depth map fusion using high-definition LiDAR data.
• Performance analysis on the KITTI benchmark dataset.

Main Results:

• All evaluated aspects (multicue, multimodality, MV classifier) individually improve accuracy.
• Fusion of visible spectrum (RGB) and depth information provides a substantial accuracy enhancement.
• The developed detector achieves top performance on the KITTI benchmark.

Conclusions:

• The fusion of RGB and depth data is a highly effective strategy for boosting object detection accuracy.
• The proposed detector, built on simple and efficient blocks, achieves state-of-the-art results.
• This approach offers a practical and computationally efficient solution for challenging real-world object detection tasks.