A Pilot Study on Intelligent Surveillance System based on Fashionpedia Dataset

This study carries out the performance evaluation of fashion attribute recognition for smart surveillance. Fashion attribute recognition aims to identify objects within an image and precisely segment clothing based on the pre-defined categories. This task holds significance as it tackles two challenges that emerge when attempting to identify subjects through direct facial recognition. The initial issue concerns the unsatisfactory performance of facial recognition when applied to CCTV data, and the second concern is related to the ethical dilemmas within society linked to using facial recognition with CCTV data. To address these difficulties, we investigated the performance of fashion attribute recognition using an object recognition model and confirmed the capability to detect and segment attributes. Further, we achieved an average accuracy of over 30% with the YOLOv8 on CCTV data. The findings imply that the network enables the identification of an object located 30 meters away from the CCTV depending on the CCTV cameras' various height and tile angle conditions. The observation can establish the foundation of advancements in smart CCTV surveillance and contribute to the intellectualization of CCTV systems.