Enhancing Power Wheelchair Mobility with a Cost-Effective Vision-Only Deep Learning System for Obstacle Detection and Avoidance

Abstract

As the demand for power wheelchair (PWC) users increases, obstacle avoidance remains a significant challenge—often beyond the capabilities of manual control. A vision-only autonomous obstacle detection and avoidance system for PWCs is introduced, using just a single monocular camera. This low-cost solution integrates deep learning, computer vision, and mobile robotics without relying on expensive depth sensors. By adapting a deep-learning model through transfer learning on a newly developed sidewalk dataset, the system effectively detects obstacles. Once detected, a novel image-space avoidance method dynamically adjusts the wheelchair's motor speeds to navigate safely around obstacles. Deployed on a standard PWC, the system demonstrates object detection accuracy comparable to conventional notebook-based methods while operating at 5 Frames Per Second (FPS) on a Raspberry Pi 4, with a total system cost under $300. Real-world tests on a university campus confirm successful detection and avoidance of nine common sidewalk obstacle classes, with the lightweight design using minimal computational resources. This shows that effective obstacle avoidance can be achieved with minimal hardware, avoiding bulky and costly sensor arrays. The proposed system offers a novel, affordable approach for retrofitting any PWC with minimal modifications using off-the-shelf components. This innovation enhances mobility for PWC users both indoors and outdoors, setting the stage for fully autonomous, vision-only PWCs in the future.

DOI: 10.61416/ceai.v28i2.9712