Active big data-based weapon risk detection and analysis and verification of violent videos - Application of Deep Learning Models and Experimental Validation for Analyzing the Risk of Weapon-Related Violence -
Korea Safety Culture Society, Tsagaantsooj Batzaya Tsagaantsooj Batzaya, Kwanghoon Pio Kim · Forum of Public Safety and Culture · 2025
Purpose: This study proposes a novel approach to detect and analyze violent crime situations involving bladed weapons in real-time, utilizing the YOLOv8m model and the CLIP concept. The aim is to overcome the limitations of existing CCTV systems and develop a system capable of predicting potential crimes in advance. Method: The YOLOv8m model was used to analyze video data, and active data was generated by storing the results in JSON format. Frames where bladed weapons and people were simultaneously detected were defined as “dangerous frames.” The CLIP concept was applied to analyze surrounding frames in intervals of 3, 5, 7, and 10 seconds. The performance was evaluated using True Positive (TP) and False Positive (FP) ratios. Results: Dangerous Object and Dangerous Frame Analysis: This study aimed to detect violent crimes involving bladed weapons in video data. The selected bladed weapon objects were a knife, baseball bat, and scissors. The YOLOv8m model was used to detect objects, and frames where a person and a weapon were detected together were defined as dangerous frames. The analysis revealed that the bladed objects were detected 646 times across 592 unique frames, and in 570 of those frames, they were detected together with a person. CLIP Concept Setting and Dangerous Frame Extraction: Out of 174,372 frames, 570 were classified as dangerous frames. The CLIP concept was applied to analyze the frequency of surrounding frames within 3, 5, 7, and 10-second intervals. CLIP Concept Performance Evaluation: Dangerous frames were labeled as “risk” or “non-risk,” and CLIP performance was evaluated. The highest performance, with an accuracy of 79.83%, was achieved in the 3-second interval, while the accuracy decreased to 62.01% in the 10-second interval. A trend of decreasing accuracy as the time range increased was observed, highlighting the importance of shorter time intervals for efficient extraction. Conclusion: Dangerous frames, where both bladed weapons and people were detected, were classified, and the performance of CLIP at different time intervals was analyzed. The detection performance was evaluated with a focus on the main bladed weapon objects (knife, scissors, and baseball bat), comparing TP and FP ratios. The 3-second interval was found to be effective in minimizing FP while maintaining TP. The study confirmed that bladed objects are closely associated with violent crimes and can be utilized in crime prevention and warning systems. Future research is suggested to include various objects and environmental variables.