Sistema de detección temprana de eventos delictivos en entidades comerciales con visión artificial y deep learning

Abstract

Presently, video surveillance systems in commercial entities depend considerably on human monitoring, which results in increased costs for these entities. Consequently, a system for early detection of criminal events using computer vision and deep learning has been proposed. This system employs two complementary processing methods: the first utilizes YOLOv8 for suspicious object detection, and the second employs an SVM algorithm that classifies key points extracted with MediaPipe for threat posture detection. The system operates on an Nvidia Jetson Nano module, which processes videos in real time and displays them in an information management system developed with Flask. It also stores detections in a SQLite database, which continuously feeds into a newly automatically tagged dataset for future model updates, and is capable of displaying historical records of detections. The system's efficacy was evaluated across three key aspects: detection of suspicious objects, with an f1-score of 87.5%; detection of threatening postures, with 74%; and early detection, with 78%.The study's findings underscore the significance of processing complementary parts, facilitating the establishment of a more comprehensive contextual understanding of the scene. The quality and breadth of the training dataset are foundational for the success or failure of computer vision and machine learning models, particularly in the context of object detection, where a diverse array of images of varying sizes, shapes, and perspectives of the object to be detected is essential for the model to generalize correctly. In the case of body postures, support vector machine (SVM) models demonstrate efficacy, though they are constrained in their ability to establish spatio-temporal contexts.