Deep Learning Obstacle Detection and Avoidance for Powered Wheelchair

Abstract

Depth sensors like RGB-D cameras, LiDARs and laser scanners are widely investigated in research for Smart Wheelchair (SW) to carry out navigation, localization and ob-stacle detection and avoidance tasks. These sensors are costly compared to monocular camera sensor. A single off-the-shelf camera can be an economically efficient sensor to achieve obstacle detection and avoidance. We present in this paper a single camera based obstacle detection and avoidance method without using any 3D information. It is a novel vision-only system for wheelchair obstacle detection and avoidance that uses a Raspberry Pi along with Raspberry Pi camera. The obstacles are detected using a deep learning model built on MobileNetV2 SSD. The model is retrained using a dedicated dataset that was built for this purpose. Bounding boxes are used to mark detected obstacles; and feed them as features to the image space obstacle avoidance module. Figure 1 depicts internal view of what does the system see and an abstract description of our system's functionality. © 2022 IEEE.