Uncertainty-aware Sidewalk Detection and Traversability for Autonomous Delivery Robots

摘要

Outdoor last-mile delivery robots operate within shared sidewalk spaces, navigating complex urban environments alongside pedestrians. These robots require the ability to understand and interpret their surroundings for successful navigation, often relying on consumer-grade RGB-D cameras. The current perception and navigation software can pose challenges, especially regarding sensing noise, environmental factors, and varying outdoor lighting conditions. The paper presents a novel approach to sidewalk detection and traversability for an autonomous delivery robot operating in and around urban environments. The proposed methodology creates a sidewalk map, using an image learning-based classifier for terrain traversability estimation. A learning-based method is used to infer terrain categories, such as sidewalks, while a geometric approach estimates properties like slope and roughness. By fusing these methods, a comprehensive obstacle map is created, encompassing various terrain attributes. This map is further employed to produce a cost map, marking areas as occupied cells to assist in path planning and obstacle avoidance. The proposed method has been evaluated via field tests conducted at different times. Results show that while considering learning-based uncertainty (semantic and traversability), both modalities complement each other compared to the conventional probabilistic traversability approaches, demonstrating a significant improvement in sidewalk detection and traversable terrain estimation.