Suspicious Activity Detection Using YOLO Pose Estimation and Classification

Abstract

     Due to the significant advancements in computer vision and the increasing incidence of theft in commercial shopping environments, the demand for automated and computerized systems capable of analyzing and detecting suspicious behaviors is paramount. Since the capability of human eyes is inconsistent in providing constant monitoring for commercial stores, a novel automated model for detecting larceny occurring in commercial stores is proposed. This proposed system focuses on the individual's circumscribed box and the body's primary key points as fundamental features trained by the  Gradient Boosting ( ) algorithm, which is better than previously designed methodologies based on video-derived training images or the spatial domain of pixel values. In this method, an augmentation of both the University of Central Florida (UCF)-established datasets and the local dataset enhanced the obtained results by exhibiting exceptional accuracy enhancement in classifying anomalous events. The accuracy levels ranged from 94.83 to 98.52% across various folds; the F1 score was 98.36%.