e-VISER

Visual servoing techniques allow to guide a robot with the use of cameras as the sole sensors. The principle consists in acquiring, processing and analyzing digital images, frame by frame, with the aim of extracting features to produce data that will be understandable by the robot, in order to take appropriate decisions for its motion. This field significantly contributes to the development of autonomous robots. However, new methods are mandatory because classical techniques are too resource consuming and barely exploit the dynamic characteristics of the perceived scene. The e-VISER project aims to design an original architecture of visual servoing, that rethinks all the data acquisition and processing chain from perception to decision, based on the event-based paradigm. This novel concept devises anew the way to measure, communicate, compute, control or actuate, only when a significant change has occurred in the dynamic of the robot or in its environment (instead of applying a constant and periodic update rate). More precisely, the idea is to combine innovative event-based elements together (typically vision sensor and controller) and develop new image processing and control strategies. Among event-based sensors, the so-called dynamic vision sensor (DVS) depicts an innovative approach. Whereas a standard camera periodically sends superfluous and redundant data, addressing all pixels for each image, a DVS only sends the pixel-level changes caused by a movement in the scene, at the time they occur. Thus, each pixel produces an event whenever it perceives a change of intensity. Therefore, DVS offers a drastic decrease in latency for perception, but its spatial resolution still remains low and far from the accuracy required for visual servoing. To overcome this weakness, it is proposed to associate both i) a DVS, to rapidly detect a movement in the scene, and ii) a high-resolution camera, to be more accurate when extracting features from the environment but with a reduced image processing (only applied to the region of interest pinpointed by the DVS events, instead of handling the entire image). This framework involves developing advanced methods for image processing, that should take into account spatial and temporal data from the DVS. The event-based controller also relies on an innovative concept: control is not computed nor updated unless it is required. Typically, triggering events occur either when a movement (or an obstacle) is detected in the scene perceived through the visual sensor, or when the control signal has to be updated in order for the robot to reach a given objective. This setup involves developing methods for image processing, control and communication in an unconventional way, while still guaranteeing stability, quality of service and performance of the closed-loop system even if the measurement and/or the control signal remain constant between two successive (non-periodic) events. To summarize, the new event-based visual servoing scheme will allow to i) reduce the computing and communication costs for image processing and control, while keeping the same accuracy and closed-loop performance, ii) speed up the visual servoing and so improve the robot performance, and iii) save energy. The whole framework will be implemented for stabilizing and flying at high speed a complex robotic system under resource constraints: an autonomous quadcopter that embeds on board the vision sensors and a computing platform for image processing and control. The ambitious experimental validation will consist of navigation and obstacle avoidance of the drone in an unstructured and unknown environment.