Non-Intrusive Head Movement Control for Powered Wheelchairs: A Vision-Based Approach

Joystick is the most common wheelchair controlling interface, however, it might not be applicable for cases of severely disabled people, for example quadriplegic users. The few solutions currently available on the market, such as the four switches on the headrest, the sip-n-puff, or the tongue drive system, can be cumbersome to use, while they offer limited control to the user. For such cases, a vision-based head-controlled wheelchair is a unique alternative solution, which despite its benefits, has not been equally well studied. In this work, we propose and evaluate a novel method for operating an Electric Powered Wheelchair (EPW) via head movements in a non-invasive way. This is based on computer vision techniques and Deep Learning to estimate accurately the orientation of the user’s head. It allows calibration for individual users and driving in a continuous navigation space as opposed to the discrete commands imposed by alternatives such as head switches. Our approach enables the design of an efficient and cost-effective solution utilising a simple RGB camera that captures the user’s face orientation. Our system is implemented and tested real-time on an EPW using its existing commercial controller, while it can work with any commercial controller the manufacturer allows interfacing with (i.e., a direct plugin). Performance is evaluated through trials conducted with healthy participants. The results (96% successful track completion) show that our head driving system can be reliably used as an alternative solution to the conventional joystick interface, with only a small trade-off in travelling time and distance (reduction by 9.4% and 21% respectively). The participants’ experience in terms of mental and physical load, subjectively assessed following the trials, suggests relatively low mental and physical demand imposed by our system. Users also expressed high confidence in the system’s performance indicating trust to the safety aspects of our implementation. Analysis of our findings and experimental observations provide a new knowledge base for potential system improvements and future designs.