Authors (including presenting author) :
Tsoi KM(1), Yu JKP(1), Chan MKL(1), Chan BCC(1), Ho AH(1), Chan CYC(1), Chan SKK(1), Ma EWL(1)
Affiliation :
(1)Hospital Authority Community Rehabilitation Service Support Center (CRSSC)
Keyword 1: :
rehabilitation
Keyword 2: :
communication
Keyword 3: :
computer vision
Keyword 4: :
neuromuscular disease
Keyword 5: :
alternative control
Introduction :
People with physical disabilities, particularly those affected by rare diseases, are restricted to interact with and manage their surroundings due to lack of suitable tools and technologies. As the result, there was difficulty in daily tasks and social activities, leading to frustration, isolation, and a diminished sense of independence.
Objectives :
This study intended to 1) investigate the effectiveness of facial expression recognition systems; 2) explore the clinical applications of system.
Methodology :
This project aims to develop a facial-expression recognition system that empowers people with physical disabilities to make choices independently. Running a computer vision model on an embedded computing board, the system detects eye-closed and tongue-out to trigger electronic components, enabling both communication and smart-home control. Volunteers with neuromuscular disorders (N=16) took part in the study. They performed two tasks: (a) activate a call bell ten times by eye-closed and tongue-out; and (b) use tongue-out, in sequence, to select five items in an augmentative and alternative communication (AAC) mobile app. Afterwards, participants rated their experience with the System Usability Scale (SUS) to gauge their willingness to use the proposed system.
Result & Outcome :
In Experiment 1, the call bell was triggered an average of 8.29 times out of 10 with tongue-out (SD = 1.38) and 9.00 times out of 10 with eye-closed (SD = 2.07). During Experiment 2, participants needed 5.78 attempts on average (SD = 0.55) to select all five items. In user acceptance assessment, a mean score of 68.9 (SD = 18.2) was resulted in the System Usability Scale. The performance of detection model was showed effective in both Experiment 1 (>80% detection accuracy) and Experiment 2 (0.78 of extra attempts needed to 5-item selection). Powered by the trained YOLOv4-tiny computer-vision model, the system lets users trigger external devices with a single facial gesture. Compared with commercial options such as eye-tracking controllers, it is less expensive and easier to set up. An average SUS score of 68.9 shows that participants judged the system’s overall performance to be better than that of more than 50% of similar products.
