Drone-Based Activity Recognition via Dynamic Features and Deep Learning
Yawar Abbas, Ahmad Jalal
Abstract
Computer vision technology has improved to the point where it has led to the development of sophisticated algorithms that are able to reliably recognize human actions from RGB videos that have been acquired by drone cameras. In spite of the fact that it has a great deal of potential, human action recognition is confronted with a multitude of obstacles, such as variances in action execution, restricted camera views, and field of view limits. The purpose of this research study is to offer a system that addresses these issues by employing RGB films obtained by drone cameras as input. The first step in the methodology involves breaking the video down into individual frames, which is then followed by the application of preprocessing techniques. Resizing the photos, turning them to grayscale in order to lessen the computing load, and removing background noise are all strategies that are utilized in these techniques. Subsequently, we use YOLOv7 algorithm for recognize human's activity from the photos and extract the skeleton points of those individuals. At this stage, we proceed to extract a variety of characteristics, such as joint angles, histograms of oriented gradients (HOG), and three-dimensional point clouds. These features are optimized through the use of stochastic gradient descent (SGD), and action categorization is carried out through the utilization of a deep learning assembler. Experiments are carried out with the Drone Action dataset and the UCF-ARG dataset in order to validate the approach. The results of these experiments demonstrate that the approach is capable of achieving notable action recognition accuracies of 85% and 75%, respectively.