Real-time Pedestrian Crossing Recognition for Assistive Outdoor Navigation

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Assistive Technology

Background:

• Urban navigation presents significant challenges for individuals with visual impairments.
• Existing navigation aids often lack robust pedestrian crossing detection capabilities.

Purpose of the Study:

• To develop and evaluate a system for real-time recognition of pedestrian crossings in outdoor urban environments.
• To provide reliable navigation cues for safe street crossing and guidance to safety islands or sidewalks.

Main Methods:

• Utilized a state-of-the-art Multisense S7S sensor to collect 3D pointcloud data.
• Developed algorithms for real-time pedestrian crossing detection and directional guidance generation.
• Integrated 3D spatial prior information from pointcloud data to improve a baseline monocular-camera system.

Main Results:

• Demonstrated improved performance over monocular-camera systems by incorporating 3D spatial data.
• Achieved robustness to occlusion and perspective transformations through adaptable system parameters.
• The system showed particular effectiveness in non-occluded scenarios and reasonable accuracy across various conditions.

Conclusions:

• The developed system offers a promising solution for enhancing urban navigation safety for the visually impaired.
• The integration of 3D pointcloud data significantly improves pedestrian crossing detection accuracy.
• A publicly available dataset of diverse pedestrian crossing scenarios supports future research in this domain.