Mitigating latency problems in vision-based autonomous UAVs

In this paper, the latency problem of computer vision systems is addressed in the framework of autonomous Unmanned Aerial Vehicles. Recent advancements in sensors and embedded electronic boards made it possible to load, even on small size drones, cameras and image processing devices. Here, a navigation system based on computer vision is considered as one of the most popular applications exploiting this technology in substitution of Global Navigation Satellite System solutions. The main issues when working with a video stream are the limited frame rate (i.e., small sampling frequency), and the non negligible computational time for extracting features from the images (i.e., latency). In particular, the latency negatively affects a position controller that exploits data from the computer vision system, preventing its usage for precise positioning applications. In this paper, a possible solution is designed according to this recipe: First, a sensor fusion technique able to compensate the latency is adopted to estimate the velocity using the position of the computer vision system and the accelerations provided by a Inertial Measurement Unit. Then, a controller is developed using two feedback loops, the inner one accounting for the estimated velocity, and the outer one exploiting the delayed position. Test experiments, showing very positive results, are finally reported.