Sistema automatizado de identificación y clasificación de fotografías de ciclistas en competencias deportivas de descenso utilizando técnicas de visión por computadora.

Abstract

Manual classification of downhill cycling event photographs by bib number and equipment color takes between 2 and 4 days per event in single-operator businesses such as Titan TV, leading to buyer abandonment and constraining business scalability. This work develops an automated computer vision system that reduces that time. Model selection combined the bibliographic review of 30 prior studies, a comparative benchmark of 16 systems over 67 reserved images and 1,290 human judgments, and training on a proprietary dataset of 1,200 photographs from 6 events. The selected models are YOLO11m for rider detection, PARSeq for optical bib recognition and Gemini 2.5 Flash for chromatic analysis. The system is deployed with a NestJS backend, a FastAPI service, a Vue 3 SPA and Cloudflare workers that serve public photographs with watermark and QR code. Operational validation over 4 events processed 2,673 photographs with zero pipeline failures. Automated bib reading covered 85.8 % of the set, with human correction limited to 4.4 % of the total. Calendar time per event dropped from about 96 hours to a range of 60 to 111 minutes, a reduction close to 98 %. Operating cost stood at 23.91 USD per month. The system guarantees complete same-day event delivery and enables scaling of sports photography production without manual work proportional to volume.