Intelligent Video Surveillance for Weapon Detection: A Novel Approach for Low-Cost, Effective Monitoring in Low-Light Environments

These days, no public safety system is complete without security and video surveillance cameras. The systems still need human intervention in many places to identify potentially dangerous scenarios. Delays in discovering occurrences or unforeseen dangers can occur when security services are understaffed, endangering the public. Using intelligence-based techniques, this project aims to create a cost-effective solution for analysing surveillance videos in real-time and detecting firearms in various scenarios. The rising annual incidence of crimes involving guns and other hazardous weapons is an issue that the police are finding increasingly difficult to address, as is evident from a number of statistics. Many areas, particularly those without gun prohibitions, have a high rate of crimes committed with knives or firearms. It is essential for public safety that crimes are detected early. Using video monitoring to identify the presence of potentially lethal weapons like knives and firearms is one approach to avoiding these kinds of incidents. Now, intervention and monitoring are also necessary for control and monitoring. Presented here is a video demonstration of a tracking system that is well-suited for controlling and monitoring targets. There has been a plethora of recent study into the use of deep learning approaches for weapon detection in video surveillance. But the detection of weapons in low-light or nighttime situations has received very little research. To fill this void, this research suggests a new method for weapon detection that is optimised for low-light settings. For this reason, an improved YOLOv7 model is used for weapon detection. We extended YOLOv7's feature extraction capabilities by integrating the CBAM attention method. To further concentrate smaller locations, we augmented the network's backbone with an SPPF+ layer. Lastly, the YOLOv7 model was updated to include decoupled heads, which allow for more efficient data extraction from arrays. More precise features were acquired and multi-scale feature fusion was accelerated with the help of a BiFPN. The use of learning weights in the BiFPN allowed the network to prioritise the characteristic mapping that had the greatest impact on the result characteristics. In order to improve the classification accuracy, the parameters of the suggested model are fine-tuned using the Modified Sooty Tern Optimisation Algorithm (MSTOA). We used a dataset of 15,367 photos for the model's training and five different dark videos to measure its performance. With a 95% accuracy score, the developed detection model does an excellent job at detecting weapons.