A Study on the Implementation of YOLOv4 Algorithm with Hyperparameter Tuning for Car Detection in Unmanned Aerial Vehicle Images

Unmanned Aerial Vehicles (UAVs) for surveillance and monitoring have become more prevalent due to their versatility and mobility. These vehicles capture high-resolution images that provide a broad field of view in real-time. Today, enhancing object detection accuracy on images captured by unmanned aerial vehicles (UAVs) has become a significant challenge. Through extensive research, it has been established that the correct setup of hyperparameters is imperative to achieving the highest accuracy in machine learning. Our study introduces a technique that utilizes hyperparameter tuning to implement the YOLOv4 algorithm, enabling the detection of cars in unmanned aerial vehicle images. In general, all scenarios of this study have different accuracy results, which have implications for their detectability. Thus, scenario 3 of YOLOv4 hyperparameter tuning is the best model accuracy. Our approach utilizes the PSU Aerial Car Images Dataset from previous studies. During this research, accuracy values were obtained through testing at the model validation stage rather than at the testing stage. In this study, we achieved a validation performance of the detection model by using a validation dataset proportion of 20%. Based on our research, it has been revealed that the YOLOv4 algorithm is a highly efficient car detection system when it comes to unmanned aerial vehicle images. Through rigorous testing of multiple hyperparameter tuning scenarios, we achieved an exceptional accuracy of 99.02% in the optimal model scenario, which utilized YOLOv4. Similarly, in replicating a research paper's hyperparameter tuning methods on YOLOv3, the highest accuracy of 98.40% was attained in scenario 2.