Drone-Based Video Surveillance Using Yolov6 and Neuro Fuzzy Classifier

The rise of computer vision technology has seen development of advanced algorithms that can accurately detect human activities from RGB videos taken by drone cameras. One major limit is the fact that human action recognition technology comes with several obstacles, for example, variations in action execution, limited camera views and field of view constraints. In this study, we intend to develop a system that addresses these challenges using RGB videos captured by drone cameras. The first step is to break down the video into frames and then apply some preprocessing techniques on each frame. These techniques include image resizing, changing it to gray scale so that it consumes less processing power and noise elimination. Subsequently, we employ YOLO algorithm for human detection and extracting their skeleton points from images. At this point various features are extracted such as joint angles, histograms of oriented gradients (HOG), and three-dimensional point clouds. These features are optimized using quadratic discriminant analysis (QDA), and action classification is performed using a Neuro-Fuzzy Classifier (NFC). In order to evaluate the approach, experiments were conducted using both Okutama Action dataset as well as Drone Action dataset. The results indicate that this method can achieve considerable accuracy in action recognition at 92% and 85% respectively.