Aplicação de visão computacional em câmeras de segurança: identificação de pessoas e objetos - detecção de níveis de perigo

Abstract

The detection of object movement through surveillance cameras has proven to be an increasingly relevant tool in various fields, such as public security, healthcare, environmental monitoring, and remote surveillance. This technology has transformed the way environments are protected and how rapid responses to risk situations are implemented. The development of autonomous algorithms combined with Deep Learning techniques—such as Convolutional Neural Networks (CNNs) and the You Only Look Once (YOLO) model—has enabled the accurate identification of weapons, violent acts, and other suspicious behaviors in real time. These systems reduce the workload on human operators and ensure greater efficiency in security and monitoring operations. However, the application of object detection techniques may lead to uncertainties and inaccuracies, such as false positives (identifying non-existent objects), false negatives (failing to detect existing objects), visual interferences, and environmental variations. These factors can compromise the reliability of detections and lead to failures in the process. One way to address these uncertainties and inaccuracies is by adjusting the threshold, which was finetuned until reaching a value of 0.47, allowing the system to maintain a high level of precision (89.85%). Additionally, the application of fuzzy logic, through the analysis of membership functions—evaluating object speed and distance—enabled a significant improvement in the project’s final accuracy, as the system was able to interpret and generate alerts only when necessary.