Synthetic Dataset Generation for Edge Drone Inference and Training

This research investigates the generation of synthetic aerial images for training deep neural networks., particularly targeting the task of object detection in mostly top-down views. The experiment leveraged a simulation environment to create a complex dataset., addressing the challenges associated with collecting real-world aerial images. The work encompassed several critical components., including algorithm parameter variation for ground truth annotation., the exploration of bounding box generation techniques and an investigation into camera orientations and altitude to simulate real-world scenarios. Additionally., the research delved into the training of neural network models., specifically tailored for low-power devices., particularly the Coral Edge TPU. The findings of this research indicate the feasibility of utilizing synthetic environments and datasets to train neural networks for aerial image analysis. The trained model showcased promising results., with the potential for broader applications in object detection. The study also addressed safety considerations., emphasizing the need for high-resolution cameras in drone operations to detect objects., such as human bodies., from safe distances. Furthermore., the paper outlines a roadmap for future work., highlighting areas such as multi-class object detection., integration with autopilot systems., and real-world testing. This research serves as a foundational step toward harnessing synthetic data and deep learning for enhanced drone capabilities in diverse scenarios., with a commitment to safety and compliance with aviation regulations.