Weapon Detection in Surveillance Videos Using YOLOV8 and PELSF-DCNN

Abstract

Weapon detection (WD) provides early detection of potentially violent situations. Despite deep learning (DL) algorithms and sophisticated closed-circuit television (CCTVs), detecting weapons is still a difficult task. So, this paper proposes a WD model using PELSF-DCNN. Initially, the input video is converted into frames and pre-processed. The objects in the pre-processed frames are detected using the YOLOv8. In meantime, motion estimation is done using the DS algorithm in the pre-processed images to cover all the information. Then, the detected weapons undergo a sliding window process by considering the motion estimated frames. The silhouette score is calculated for detected humans and other objects. Now, the features are extracted and the important features are selected using the CSBO algorithm. The selected features and the output of YOLOv8 are given to the PELSF-DCNN classifier. Finally, the confidence score is calculated for the frame to define the number of weapons. In an experimental evaluation, the proposed method is found to be more efficient than the existing methods.