In recent years, UAV Earth Observation (EO) generated a vast amount of data, enabling vital problems to be solved. To keep pace with this exponential growth, new data processing solutions are needed. This includes not only artificial intelligence processing, but also the need for real-time processing. Object detection is one of the main purposes of computer vision (CV), with increasing demand for deployment on resource-constrained embedded systems. This paper investigates the possibilities of real-time Edge AI processing of visual aerial data (images and videos) on such platforms using deep learning (DL) techniques. We compared a range of state-of-the-art lightweight discriminative convolutional neural network (CNN) architectures, including MobileNet, EfficientNet and YOLO variants, assessing their performance against key metrics such as accuracy and inference speed. In addition, we explored the feasibility of hardware optimization and acceleration using embedded GPUs and specialized neural network accelerators to further enhance real-time capabilities. This comprehensive benchmark provides valuable information for developers seeking to optimize object recognition models for deployment in resource-constrained embedded systems, thus contributing to the advancement of intelligent vision applications in fields such as robotics and surveillance.