Recent Trends of Resilient UAV On-Board Vision Processing

The number, variety, and applications of Unmanned Air Vehicles (UAVs), have grown exponentially in the last few years, and is expected to continue in the future. Modern UAVs should be highly accurate, available, scalable, and cost-effective as well as consider power efficiency and computational complexity for on-board vision processing. In this paper, the related work of recent trends of resilient UAV on-board vision processing focused on the fields of image and video processing is described. Then, the results from two experimental implementations are presented associated with: (i) object detection and tracking and, (ii) visual-based navigation. For the first case, efficient pre-trained deep learning object detection models and robust trackers were deployed on an edge computing device. Then an evaluation of the selected object detection models and trackers was implemented to compare them in terms of the related metrics, inference time and computing resources. For the second case, an experimental implementation of a visual-based autonomous navigation system was implemented with the capability to be deployed for UAV for on-board processing. A multi-sensor approach based on Simultaneous Localization and Mapping (SLAM) processing was adopted. Then, a qualitative validation of the on-board processing efficiency and navigation accuracy in outdoor environments was carried out. For both cases, the results highlight the capability of the adopted approaches to be deployed for on-board processing in real-time terms.