Cross-Safe: A Computer Vision-Based Approach to Make All Intersection-Related Pedestrian Signals Accessible for the Visually Impaired

摘要

Intersections pose great challenges to blind or visually impaired travelers who aim to cross roads safely and efficiently given unpredictable traffic control. Due to decreases in vision and increasingly difficult odds when planning and negotiating dynamic environments, visually impaired travelers require devices and/or assistance (i.e. cane, talking signals) to successfully execute intersection navigation. The proposed research project is to develop a novel computer vision-based approach, named Cross-Safe, that provides accurate and accessible guidance to the visually impaired as one crosses intersections, as part of a larger unified smart wearable device. As a first step, we focused on the red-light-green-light, go-no-go problem, as accessible pedestrian signals are drastically missing from urban infrastructure in New York City. Cross-Safe leverages state-of-the-art deep learning techniques for realtime pedestrian signal detection and recognition. A portable GPU unit, the Nvidia Jetson TX2, provides mobile visual computing and a cognitive assistant provides accurate voice-based guidance. More specifically, a lighter recognition algorithm was developed and equipped for Cross-Safe, enabling robust walking signal sign detection and signal recognition. Recognized signals are conveyed to visually impaired end user by vocal guidance, providing critical information for real-time intersection navigation. Cross-Safe is also able to balance portability, recognition accuracy, computing efficiency and power consumption. A custom image library was built and developed to train, validate, and test our methodology on real traffic intersections, demonstrating the feasibility of Cross-Safe in providing safe guidance to the visually impaired at urban intersections. Subsequently, experimental results show robust preliminary findings of our detection and recognition algorithm.