Lightweight Anomaly Detection Mechanism Based on Machine Learning Using Low-Cost Surveillance Cameras

As the need for on-site monitoring using surveillance cameras increases, there has been a growing interest in automation research incorporating machine learning. However, traditional research has not resolved the performance and resource efficiency trade-offs. Traditional research often utilizes high-resolution images to enhance detection performance. However, surveillance cameras, being Internet of Things devices, are constrained by limited resources, making high-resolution images less suitable for their operation. Therefore, we proposed a lightweight learning model that is more efficient and with minimal performance degradation. The proposed model reduces the resolution of the image until the performance is maintained, finding where the trade-off is resolved for each dataset. It is also utilized for real-time detection by determining the probability values of detection at an appropriate resolution. Using this, we suggested a real-time lightweight fire detection algorithm. The proposed mechanism is approximately 30 times more memory efficient while maintaining the detection performance of traditional methods.