Handheld Knife Stick Detection Based on Dual-Path Multi-layer Residuals

Due to strict gun control policies, many countries completely prohibit private firearm ownership, leading to personal conflicts often involving handheld knives and sticks as the primary weapons. Detection of handheld knives and sticks in specific scenarios can effectively prevent conflicts and accidents. However, handheld knife and stick detection faces challenges such as limited datasets, severe occlusions, diverse types of knives and sticks, and difficulties in detecting slender objects, all of which significantly affect detection accuracy. This paper addresses the issue of limited datasets by creating a Handheld Knife Stick Detection (HKSD) dataset consisting of 1850 images from various scenes, including some from publicly available datasets. Additionally, to tackle challenges such as severe occlusions and diverse shapes, a detection model called DMR-YOLO is proposed in this study. DMR-YOLO addresses severe occlusion by introducing a Dual-path Multi-layer Residual Backbone (DMRB) network, which retains more original features to enhance the model’s selection of useful information. Furthermore, DMR-YOLO incorporates the concept of dual-path multi-layer residuals into a feature pyramid, proposing a novel Dual-Path Feature Pyramid (DPFP) to enhance the model’s robustness in extracting semantic information from various knife and stick features. To address detection challenges posed by excessively slender objects, DMR-YOLO introduces a module based on Transformer Architecture Fused with Soft Attention mechanisms (TAFSA). TAFSA provides global attention to the model, thereby improving the detection accuracy of slender knives and sticks. Experimental results demonstrate that the proposed DMR-YOLO model surpasses existing state-of-the-art detection models in terms of detection accuracy.