Search

International audience; Artificial Intelligence and Decision Making Challenges in the Certification of Computer Vision-Based Systems for Civil Aeronautics C omputer vision techniques have made considerable progress in recent years. This advance now makes possible the practical use of computer vision in civil drones or aircraft, replacing human pilots. The question that naturally arises is then to provide a way to certify those types of systems at a given level of safety. The aim of the article is, firstly, to understand the gap between today's computer vision systems and the current certification standards; secondly, to identify the key activities that must be fulfilled in order to make computer-vision systems certifiable and, thirdly, to explore some recent works related to these key activities.

Artificial Intelligence (AI) is currently an inescapable keyword in computer science given its predicted huge contribution to the global economy [1] or to the whole society [2, 3], as argued in many recent white papers or reports. Its techniques are expected to provide efficient ways to deal with heterogeneous and voluminous numerical environments and data, help or even improve decision making, and automate complex functions. The aeronautics, spatial and defense domain (ASD) is impacted by this evolution [4]. AI is a research field with a complex history and numerous areas. It is customary to divide these into two trends: Formal and logical, which rely on models, knowledge representation and solvers; Empirical, which rely on data, statistical estimation and inference. Although this last trend mainly drives what is sometimes referred to as the third wave of AI, with Machine Learning playing the key role of a general design principle, it cannot fully solve all problems: data can be rare and costly in the ASD domain, where the requirements of reliability and predictability are often very high. The present issue of the Aerospace Lab journal contains several examples of AI research, from rather specific studies to more general position papers or surveys, which exemplifies these two traditions and, eventually, their possible combination., AerospaceLab Journal, Issue 15, September 2020; ISSN: 2107-6596

International audience; Tractable algorithms used for 6DOF visual-inertial odometry have decades-long history of estimation consistency issues. Those arise in particular in two well-studied filters: namely the EKF-SLAM and MSCKF. Recently, strong theoretical works linked the error-state of these filters with their consistency properties; these results led to the synthesis of far more consistent filters. In previous works, we have shown that using similar filter for the fusion of magneto-inertial sensors with optical ones improved classical visual-inertial navigation systems. The consistency of those novel magneto-visual-inertial filters were, however, not addressed until now. This work does. We apply invariance theory findings to the specific case of magneto-inertial odometry and magneto-visual-inertial odometry for the synthesis of a filter with interesting consistency properties. We describe thoroughly such an invariant filter, implement it and conduct experiments on carefully captured data from real sensor. By comparing the results of non-invariant, observability-constrained and invariant versions of the filter, we find that the invariant version (i) shows an error estimate that is consistent with observability of the system, (ii) is applicable in case of unknown heading at initialization, (iii) improves long-term behavior of the filter and (iv) exhibits a lower normalized estimation error. We experiment on challenging scenarios for regular visual-inertial pedestrian navigation systems.

International audience; Visual-inertial Navigation Systems (VINS) are nowadays used for robotic or augmented reality applications. They aim to compute the motion of the robot or the pedestrian in an environment that is unknown and does not have specific localization infrastructure. Because of the low quality of inertial sensors that can be used reasonably for these two applications, state of the art VINS rely heavily on the visual information to correct at high frequency the drift of inertial sensors integration. These methods struggle when environment does not provide usable visual features, such than in low-light of texture-less areas. In the last few years, some work have been focused on using an array of magnetometers to exploit opportunistic stationary magnetic disturbances available indoor in order to deduce a velocity. This led to Magneto-inertial Dead-reckoning (MI-DR) systems that show interesting performance in their nominal conditions, even if they can be defeated when the local magnetic gradient is too low, for example outdoor. We propose in this work to fuse the information from a monocular camera with the MI-DR technique to increase the robustness of both traditional VINS and MI-DR itself. We use an inverse square root filter inspired by the MSCKF algorithm and describe its structure thoroughly in this paper. We show navigation results on a real dataset captured by a sensor fusing a commercial-grade camera with our custom MIMU (Magneto-inertial Measurment Unit) sensor. The fused estimate demonstrates higher robustness compared to pure VINS estimate, specially in areas where vision is non informative. These results could ultimately increase the working domain of mobile augmented reality systems.

International audience; This paper addresses the problem of global visual ego-localization of a robot equipped with a monocular camera that has to navigate autonomously in an urban environment. The robot has access to a database of geo-referenced images of its environment and to the outputs of an odometric system (Inertial Measurement Unit or visual odometry). We suppose that no GPS information is available. The goal of the approach described and evaluated in this paper is to exploit a Hidden Markov Model (HMM) to combine the localization estimates provided by the odometric system and the visual similarities between acquired images and the geo-localized image database. It is shown that the use of spatial and temporal constraints reduces the mean localization error from 16 m to 4 m over a 11 km path evaluated on the Google Pittsburgh dataset when compared to an image based method alone.

International audience; This paper presents a complete mobile robot architecture for autonomous exploration in a GPS-denied unknown environment. The platform considered is equipped with wheel encoders, stereo-vision and depth sensors, the measurements of which are fused within an extended Kalman filter for robust localization. An occupancy grid of the environment is built on-line for environment reconstruction and obstacle detection. Based on this map, a model predictive control scheme autonomously defines safe exploration trajectories, while taking into account interaction with the imaging sensors. Experimental results demonstrate the embedded computational capability of this vision-based control loop.