AI-Based Security System Using YOLO Algorithms

Abstract

Throughout the first chapter of this thesis, we have presented a comprehensive examination of security systems, addressing crucial considerations essential for designing robust solutions. We have highlighted key challenges faced by these systems, notably human errors leading to the oversight of significant events, the inherent complexity of system architecture, and the difficulties encountered during system extension and updates. This study offers an effective solution to the aforementioned challenges. By utilizing an AI card such as the Nvidia Jetson Nano, existing security camera systems can be transformed into intelligent and robust entities. This integration enhances their ability to process relevant events with high accuracy while potentially eliminating the need for additional equipment, replaced instead by the fusion of AI algorithms with the visual data captured by the cameras. The YOLOv8 model was trained using a large dataset downloaded from the Roboflow platform. Its images are labeled with seven classes: customer bagpack, null, product, product-picked, regular, shoplifting, and shopping cart. With a large configuration (43.7M parameters), we have obtained a good accuracy (0.9) and satisfactory convergence. However, during testing in different retail environments, challenges arose in accurately detecting certain products.