An Efficient Detection of Suspicious Objects from Dynamic Video Surveillance by Fusion-based Multiview Deep Learning Techniques

Abstract

Real-time detections of suspicious objects are needed to identify for finding criminal activities and are used in immediate alert systems for public safety applications. Video surveillance systems use live, closed-circuit televisions (live CCTVs) for dynamic video capturing of objects. Finding criminal activities over the dynamic video data is an emerging surveillance problem. The deep learning techniques are tedious for detecting suspicious movable objects and criminal activities. YOLO (You Only Look Once) gives more prominent movable video object detection accuracy than conventional deep models, like Convolutional Neural Network (CNN), 3D CNN, and Convolutional LSTM. State-of-the-art YOLO models, YOLOv8n, YOLOv8s, and YOLOv8l, are emphasized for extracting and detecting object motion detection from the dynamic video. YOLO models use single-view deep learning to classify or detect objects. These models limit the accuracy of the detection of complex and dynamic objects of dynamic video data. This paper presents the Fusion-based Multiview deep learning techniques to overcome this issue. The experimental study demonstrates that the proposed methodology efficiently detects suspicious data objects more than the single-view deep models.