UAV Embedded Real-Time Object Detection by a DCNN Model Trained on Synthetic Dataset

The utilization of unmanned aerial vehicles (UAVs) in civilian and military applications has significantly increased in recent years. A common task associated with these applications is detecting objects of interest in images captured by onboard UAV cameras. The ongoing development of advanced deep convolutional neural network (DCNN) algorithms has substantially improved the accuracy of general image segmentation and classification. However, applying these techniques to images obtained from UAVs requires a representative dataset for enhanced performance. This paper presents a method for DCNN-based object detection, utilizing resources embedded in a 1.5kg quadrotor-type UAV. To address the lack of representative datasets for our target scope, we employed a DCNN model trained on a self-generated synthetic dataset. The proposed method has been validated through real experiments, and the results demonstrate this approach's feasibility for real-time surveillance with fully onboard processing. Furthermore, this offers a stand-alone, portable, and cost-effective solution for surveillance tasks using a small UAV.