A Wearable Face Recognition System for Individuals with Visual Impairments
Sreekar Krishna, Greg Little, John Black, Sethuraman Panchanathan · 2005 · Proceedings of the 7th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '05) · doi:10.1145/1090785.1090806
Summary
This paper presents the iCare Interaction Assistant, a wearable face recognition system developed at Arizona State University's Center for Cognitive Ubiquitous Computing (CUbiC) to help people who are visually impaired identify individuals during social interactions. The system consists of camera-equipped glasses with a pinhole camera, a video digitiser, and a laptop running face detection and recognition algorithms, with results spoken to the user through headphones via text-to-speech. The research addresses the significant challenge that face recognition algorithms trained on controlled laboratory datasets perform poorly in real-world conditions where facial pose angles and illumination vary considerably. The authors created the FacePix(30) database — a custom dataset of 30 individuals photographed at pose angles from -90 to +90 degrees and illumination angles from -90 to +90 degrees (at 1-degree increments), producing 181 pose images and 181 illumination images per subject. Using this database, they systematically evaluated the robustness of four widely used face recognition algorithms — Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Bayesian Intrapersonal Classifier (BIC), and Hidden Markov Model (HMM) — across varying pose and illumination conditions.
Key findings
The robustness evaluation revealed that LDA performed best overall for both pose angle variation (robustness scores: 0.3485 to 0.9245 depending on training set breadth) and illumination changes (0.7820 to 0.9888), while HMM was the weakest performer. BIC performed very poorly for pose changes but reasonably well for illumination. PCA performed similarly to LDA for pose but less well for illumination. When tested on images captured from the actual wearable device in an office environment (450 images of 10 individuals), PCA achieved a higher recognition rate than LDA despite LDA being twice as fast — so PCA was selected for the wearable device. The system required recognition of the same individual in five consecutive video frames before announcing the name, to guard against the problem that changing lighting conditions in the environment could cause sporadic misidentification. Recognition times averaged 40.6ms per face for LDA and 85.8ms for PCA in the wearable device experiments.
Relevance
This early work on wearable face recognition for people who are blind anticipated a technology area that has since become commercially available through smartphone apps and smart glasses. The fundamental challenge the authors identified — that face recognition performance degrades significantly with pose and illumination changes in real-world settings — remains relevant even with modern deep learning approaches. For accessibility practitioners, this research highlights the social dimension of visual impairment that goes beyond navigation and text access: the inability to recognise people creates barriers to social participation, professional networking, and the small daily interactions that build community. The paper's approach of creating a realistic evaluation database with systematic pose and illumination variation, rather than relying on idealised lab datasets, established a methodology that helped move assistive face recognition toward practical deployment. Modern implementations (like Microsoft's Seeing AI or Be My Eyes) build on this foundation, though they benefit from vastly more powerful hardware and deep learning algorithms.
Tags: face recognition · wearable computing · blindness · computer vision · social interaction · assistive technology · visual impairment