Visão computacional embarcada para acessibilidade: otimização da rede YOLO em Raspberry Pi para detecção e classificação tipológica de cadeiras de rodas

Abstract

Accessibility in the built environment depends on a space’s ability to recognize the specific needs of the people who use it. Current building automation systems, however, treat people who use wheelchairs in the same way as everyone else and treat morphologically distinct devices as a single category, which limits the precision of environmental adaptations and undermines the potential of smart assistive technologies. This dissertation investigates to what extent the combination of a consolidated typological classification and the optimization of detection models in embedded hardware can constitute a low-cost platform capable of identifying specific types of wheelchairs in real time, providing technical support for future integration with automation systems such as doors, elevators, and smart signage. The study, which takes an applied and exploratory approach, is based on the hypothetical-deductive method in accordance with the guidelines of Marconi and Lakatos (2017) and adopts as its operational hypothesis the achievement of a mean accuracy mAP@ 0.5 of 0.85 or higher and an inference rate of 15 frames per second or higher on the Raspberry Pi, under the overarching premise of low cost and preservation of user privacy. Based on the ISO 9999:2022 and ABNT NBR 9050:2020 standards, as well as catalogs from four domestic and international manufacturers, this study establishes a classification of eleven types for the purposes of computer-aided detection. The corpus comprises 19,452 raw images sourced from four public repositories and two rounds of targeted searches, reduced, after successive filtering by resolution, deduplication, and visual screening, to 2,000 images annotated in LabelImg software and stratified into 1,394 training images, 394 validation images, and 212 test images, unevenly distributed across the eleven classes. The study trains the YOLOv5n, YOLOv5s, YOLOv8n, and YOLOv8s, of which YOLOv8n is selected for deployment due to its aggregate mAP@0.5 of 0.711; although this value does not reach the formal threshold of 0.85, four classes individually exceeded it and a fifth remained in a borderline situation. The model is converted to the NCNN format in half-precision with a drop of only 0.003 absolute points in accuracy (mAP@0.5 of 0.708) and operates on the Raspberry Pi 5 at 13.3 frames per second in a real-world scenario and 15.1 frames per second in a synthetic benchmark (so that the 15 FPS threshold was reached only under synthetic conditions), a value largely consistent with the travel speed of wheelchair users reported by Pinnock Branford et al. (2025) at 0.43 meters per second. The system integrates three applications linked through methodological triangulation, operates autonomously for four to five hours on battery power via the Geekworm X1202 UPS module, and maintains thermal stability under sustained operation. The findings highlight an architectural paradox rarely discussed in the literature, according to which, in a moderately unbalanced corpus, increasing the number of parameters can impair performance, and demonstrate the technical and economic feasibility of embedded typological detection as a perception layer for a new generation of adaptations sensitive to the diversity of mobility equipment in the built environment.