Enhancing UAV Classification With Synthetic Data: GMM LiDAR Simulator for Aerial Surveillance Applications

The proliferation of unmanned aerial vehicles (UAVs) has raised safety concerns due to the potential threats resulting from their misuse or malicious intent. Due to their compact size, high-resolution surveillance systems such as light detecting and ranging (LiDAR) sensors are necessary to exert effective control over the airspace. Given the large volume of data that these technologies generate, efficient deep learning (DL) algorithms are needed to make their real-time implementation feasible. However, the training of DL models requires extense and diverse datasets, which in certain scenarios may not be available. Therefore, this work introduces a novel method based on Gaussian mixture models (GMMs) for simulating realistic synthetic point clouds of UAVs. This simulator is calibrated using experimental data and allows to probabilistically replicate the intricacies of sensor ray propagation, thereby addressing the limitations of current ray tracing (RT) simulators such as Gazebo or CARLA. In this study, we perform a quantitative analysis of the point cloud quality of the GMM simulator, comparing it with the results obtained using a RT approach. In addition, we evaluate the effectiveness of both methods in training object classifiers. Results demonstrate the GMM simulator’s potential for creating realistic synthetic databases.