Real-time attack detection using deep learning techniques

Security has been one of the most important concepts for humans since their existence, and it has consistently increased in significance over time. In this context, the use of security cameras or the presence of personnel on duty has become a necessity to ensure both individual and public safety. However, it is not feasible for human operators to continuously monitor a security camera or a specific area with the highest level of attention throughout their shift. The need for security systems that can automatically analyze data from images with fewer human operations is increasingly growing to establish security. When considering security applications, tasks such as surveillance, target detection, and target tracking are crucial, where motion detection and classification of this motion play a significant role. The conducted study aims to achieve real-time attack detection from images obtained from a "drone" camera. Optical flow, camera motion compensation, and YOLOv8 segmentation algorithms were sequentially used for motion detection from images, while for human activity recognition, YOLOv8 segmentation and object detection algorithms were employed. Additionally, attempts were made to detect the object's distance from the "drone" using camera images. All variants of YOLOv8 (Nano, Small, Medium, Large, and Extra Large) were individually utilized in this study, and their performances were compared. The most successful YOLOv8 variant in motion detection achieved a value of 89% mAP@0.5 with YOLOv8m, while in object detection, it was 96% mAP@0.5 with YOLOv8l. The object's distance to the drone was calculated with a relative error average of 1.66%.