SIRD-YOLO: an enhanced deep learning model for weapon detection using spatial interactions and diverse receptive fields

Automated detection of weapons during video surveillance of high-risk zones is a vital requirement nowadays. Even though, a variety of deep learning models and weapon-datasets have been proposed in literature, they often encounter challenges when applied to real world scenarios, which have occlusions, cluttered backgrounds, varying luminosity conditions, etc. Besides these, the necessity to detect weapons of varying-sizes pose a different problem. To address these challenges, this paper presents SIRD-YOLO, a deep learning-based weapon detection approach that captures contextual information through spatial-interactions among the intermediate features with receptive diverse fields. Realising the importance of domain specific datasets on ML-based object detection techniques, we have developed a weapon-dataset called IITP-W, which captures scenarios reflected in the real world. The results of the performance benchmarking of SIRD-YOLO with IITP-W and other publicly available datasets showcase a significant improvement over the existing state-of-the-art models, in detecting weapons under real-world context.