Improving smart home surveillance through YOLO model with transfer learning and quantization for enhanced accuracy and efficiency

Abstract

The use of closed-circuit television (CCTV) systems is widespread in all areas where serious safety concerns exist. Keeping an eye on things manually sounds like a time-consuming and challenging process. Identifying theft, detecting aggression, detecting explosive risks, etc., are all circumstances in which the term “security” takes on multiple meanings. When applied to crowded public spaces, the phrase “security” encompasses nearly every conceivable kind of abnormality. Detecting violent behaviour among them is challenging since it typically occurs in a group setting. Several practical limitations make it hard, though complex functional limitations make it difficult to analyze crowd film scenes for anomalous or aberrant behaviour. This article provides a broad overview of the field, starting with object identification and moving on to action recognition, crowd analysis, and violence detection in a crowd setting. By combining you only look once (YOLO) with transfer learning, the model may acquire new skills from various sources. This makes it more flexible for use in various object identification applications and lessens the time and effort required to gather large annotated datasets. This article proposes the YOLO model with transfer learning for intelligent surveillance in Internet of Thing (IoT)-enabled home environments in smart cities. Quantization concepts are being applied to optimize the YOLO model in this work. Using YOLO with quantization, the model is optimized for use on edge devices and mobile platforms, which have limited computing capabilities. Thus, even with limited technology, object detection systems may be used in various real-world applications. The proposed model has been validated on two different datasets of 7,382 images. The proposed model gains an accuracy level of 98.27%. The proposed method outperforms the conventional one. The use of the YOLO model and transfer learning with quantization has significant potential for enhancing ecological smart city monitoring, and further research and development in this area could contribute to developing more effective and efficient environmental smart city monitoring systems.