A GA-Based Learning Strategy Applied to YOLOv5 for Human Object Detection in UAV Surveillance System

YOLO (You-Only-Look-Once) is a deep learning-based one-stage detection method that has been widely used and achieved great success in image classification and localization. As the state-of-the-art method, YOLO has been upgraded to version 5. This paper proposes a new approach to using a Genetic Algorithm (GA) within a YOLOv5 framework for human object detection applied in the Unmanned Aerial Vehicle (UAV) perspective image dataset. The dataset has challenges, such as a small target, the view of the object is from above, and there is an illumination and light effect. To comply with this challenge, we will utilize the dataset of visual images taken from a UAV (RGB-image) along with Thermal Infrared (TIR) information. GA is used for optimizing the Hyperparameter, which is one of the critical factors in determining the model’s performance. Based on our numerical experiments, we found that this YOLOv5-based transfer learning method using RGB-TIR dataset and optimized by GA can achieve higher accuracy compared with the original YOLOv5 for Human Detection on Unmanned Aerial Vehicle Perspective. The objective of this research is to create a surveillance system that will be used to monitor a wide area using autonomous UAVs that can exchange information with each other. In the end, the solution from this research can help related parties in tackling the problem of illegal activities with limited human resources.