• machine learning
• power wheelchair
• low cost
• assistive technologies
• image segmentation
• scene parsing
• artificial intelligence
• real time
• Jetson Nano

In modern society, an ever-growing focus and attention are given to accessibility and enabling technologies for people with disabilities. In the mobility field, people with motor skill impairments exploit power wheelchair to move in indoor and outdoor scenarios. In this thesis we want to explore the concept of an assistive power wheelchair for outdoor uses, providing obstacle detection and avoidance in urban scenarios, by leveraging low-cost digital cameras and machine learning. This work focuses on the design and development of a deep neural network for semantic segmentation of urban scenes, with the ability to recognize obstacles in an outdoor setting. Starting from an analysis of the state of the art, five networks were identified as potential starting points. Each network was trained on a dataset representative of the final environment. The selected network was optimized using the NVIDIA TensorRT SDK to perform inference at the edge on a Jetson Nano hardware accelerator, to be mounted directly on the wheelchair. The resulting model achieves an accuracy of around 85% and inference times of 35ms, thus providing a concrete solution towards the target assisted power wheelchair.