Dynamic Fusion Module Evolves Drivable Area and Road Anomaly Detection: A Benchmark and Algorithms.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Joint detection of drivable areas and road anomalies is crucial for mobile robot navigation.
• Existing benchmarks primarily cater to self-driving cars, lacking specific relevance for ground mobile robots like robotic wheelchairs.
• Convolutional Neural Networks (CNNs) are widely used for semantic segmentation in road scene understanding.

Purpose of the Study:

• To establish a dedicated benchmark for drivable area and road anomaly detection tailored for ground mobile robots.
• To evaluate the performance of state-of-the-art (SOTA) single-modal and data-fusion CNNs on this new benchmark.
• To introduce and validate a novel Dynamic Fusion Module (DFM) for enhanced visual feature fusion in existing networks.

Main Methods:

• Development of a new benchmark dataset for ground mobile robot perception.
• Evaluation of six visual feature modalities using SOTA semantic segmentation CNNs.
• Implementation and integration of the proposed Dynamic Fusion Module (DFM) into data-fusion networks.

Main Results:

• The transformed disparity image was identified as the most informative visual feature for this task.
• The proposed DFM-RTFNet significantly outperformed existing SOTA methods on the new benchmark.
• DFM-RTFNet demonstrated competitive performance on the established KITTI road benchmark.

Conclusions:

• The developed benchmark provides a valuable resource for advancing road scene understanding in ground mobile robotics.
• The Dynamic Fusion Module (DFM) offers an effective and efficient approach to fuse diverse visual features.
• The DFM-RTFNet represents a significant advancement in drivable area and road anomaly detection for mobile robots.