Detecting Threats from Live Videos using Deep Learning Algorithms

Threat detection is an important area of research, particularly in security and surveillance applications. The research is focused on developing a threat detection system using DL techniques. The system aims to detect potential threats in real-time video streams, enabling early identification and timely response to potential security risks. The study uses two state-of-the-art DL models, MobileNet and YOLOv5, to train the object detection system. The TensorFlow object detection API is employed for training and evaluating the models. The results of the study indicate that MobileNet outperforms YOLOv5 in terms of detection accuracy, speed, and overall performance. The justification for selecting MobileNet over YOLOv5 is based on several factors. First, MobileNet has a lightweight architecture, making it suitable for real-time applications where processing speed is critical. Second, it is efficient in terms of memory usage, enabling it to operate effectively on low-resource devices. Third, MobileNet provides high accuracy in detecting objects of different sizes and shapes. The study evaluated the performance of the threat detection system using various evaluation metrics, including mean average recall (mAR), mean average precision (mAP) and Intersection over union (IoU). The results show that the system achieved high accuracy in detecting threats, with an overall mAP (mean average precision) of 0.9125, mAR (mean average recall) of 0.9565 and Intersection over union (IoU) of 0.9045. In this study, researchers present a highly efficient and successful method for identifying threats through the utilization of deep learning methods. The research demonstrates the superiority of MobileNet over YOLOv5 in terms of performance, and the results obtained validate the effectiveness of the proposed system in detecting potential threats in real-time video streams.