Unmanned Aerial Vehicle for Human Surveillance using Multi-Features and Neuro-Fuzzy Classifier

Due to their bird's eye view capabilities, drones have expanded use across various domains such as drone imagery, human activity recognition (HAR), search and rescue (SAR), and surveillance. However, this advantage also presents challenges, including the dynamic nature of backgrounds and obscured perspectives. This study presents a novel approach to addressing these challenges in HAR using drone videos. The method begins with extracting frames from drone videos, followed by preprocessing steps such as gamma correction and background removal. YOLO is employed for human detection, followed by keypoint extraction to capture critical human joint locations. Comprehensive features are extracted using different feature extraction techniques. Full-body features are extracted using BRIEF and 3D point clouds, and keypoint features are extracted using spin images and USC feature descriptors. These features are then optimized using kernel Discriminant Analysis (KDA) and the neuro-fuzzy classifier to classify different actions. The proposed mode is evaluated on two benchmark datasets: UAV human and Drone Action datasets. The model achieved notable accuracies of 55.4% on the UAV human dataset and 85.5% on the Drone Action dataset.