Real-Time Image Based Weapon Detection Using YOLO Algorithms

From last few years country faces major challenge in maintaining security standards particularly in public and highly sensitive places such as airports, movie theatres, stadiums, and national parks, etc. The offsite and onsite planers use many tactics to resist authority and disrupt and to add turmoil in order to achieve their goals and objectives. These tactics can be planned or unplanned. Many crowd management experts suggest that the availability of suspicious objects such as Camera, Handgun, Rifles, Dagger, Sword and Sticks at sight before or during the event can be an indication of upcoming threat or any unlawful activities and their identification may help security forces in their proactive management and control of any destructive activities. In this research work, we generated a novel dataset "DIAT-Weapon" Dataset for weapon object detection using web scraping techniques. DIAT-Weapon Dataset consists of 2712 images divided into six categories mainly: Camera, Handgun, Rifles, Dagger, Sword and Sticks. We customized and fine-tuned YOLOv4 models to classify and position the six types of harmful objects i.e. Camera, Handgun, Rifles, Dagger, Sword, and Sticks in real time. To achieve real-time faster performance and better detection accuracy, YOLOv4 is fine-tuned, and the preset anchors trained on DIAT-Weapon annotated dataset. Using a series of YOLOv4 object detection algorithms, we demonstrated experiments on our dataset, achieving 0.63 mAP. To our best knowledge, this is the first work that utilizes customized YOLOv4 model for real-time localization and classification of weapon objects into six different categories. To our best knowledge, this is the first work that utilizes customized YOLOv4 model for real time localization and classification of weapon objects into six different categories.