Weapon Detection Method Using Convolutional Neural Network with Human Pose Estimation

The increasing incidence of violent events involving weapons highlights the urgent need for advanced technologies to prevent and forewarn individuals. Currently, closed-circuit television (CCTV) systems are widely utilized to document incidents for use as evidence; however, these systems lack the capability to prevent, detect, or notify stakeholders of incidents in real time. This study proposes the development of a new system designed to monitor and mitigate violent incidents involving weapons by proposing an innovative method to improve the accuracy of weapon detection within CCTV frameworks. The proposed method initially detects weapons within sequential images. In instances where detection fails, the system employs a human pose estimation model to reassess the potential presence of weapons. Upon identifying a detected individual, the system utilizes an adaptive pose region to refine the area of interest for further weapon detection. From experimental results, the proposed method significantly improves performance metrics, achieving average values of 92% in precision, 91% in recall, and 92% in F1-score on the testing dataset. These findings underscore the effectiveness of the proposed system in detecting and identifying weapons associated with violent incidents, thereby contributing to the prevention of future violent events.