For neuromorphic vision sensors such as event-based cameras, a paradigm shift is required to adapt optical flow estimation as it is critical for many applications. Regarding the costly computations, Principal Component Analysis (PCA) approach is adapted to the problem of event-based optical flow estimation. We propose different PCA regularization methods enhancing the optical flow estimation efficiently. Furthermore, we show that the variants of our proposed method, dedicated to real-time context, are about two times faster than state-of-the-art implementations while significantly improving optical flow accuracy.

The emergence of neuromorphic vision sensors (also known as event-based cameras) revolutionized computer vision algorithms in the field of robotics. However, the pixel-independent nature of events and the lack of photometric information raised new challenges. One of the challenges in computer vision, such as line-based SLAM (Simultaneous Localization and Mapping) and visual odometry schemes, is line detection and segmentation. In this paper, we address this problem and provide a fast method to compute line detection by avoiding time-consuming search algorithms and limiting the use of any complex implementation. Our algorithm exploits the geometrical features of the scene and applies simple heuristics to ensure correct line detection and segmentation. As a result, the inherent line detection problems that lead to either heavy computation or false estimation, like the use of search algorithms or wrong line connection, are addressed and tackled. Furthermore, the results are shown to be accurate in different motion scenarios (rotational and translational) and robust against false line segmentation.

This paper introduces a Wearable SLAM system that performs indoor and outdoor SLAM in real time. The related project is part of the MALIN challenge which aims at creating a system to track emergency response agents in complex scenarios (such as dark environments, smoked rooms, repetitive patterns, building floor transitions and doorway crossing problems), where GPS technology is insufficient or inoperative. The proposed system fuses different SLAM technologies to compensate the lack of robustness of each, while estimating the pose individually. LiDAR and visual SLAM are fused with an inertial sensor in such a way that the system is able to maintain GPS coordinates that are sent via radio to a ground station, for real-time tracking. More specifically, LiDAR and monocular vision technologies are tested in dynamic scenarios where the main advantages of each have been evaluated and compared. Finally, 3D reconstruction up to three levels of details is performed.

This paper describes an extension of the popular simultaneous localisation and mapping system (RGB-D SLAM) introduced by Endres et al. In [2]. RGB-D SLAM uses a moving RGB-D sensor (i.e. A Kinect) to incrementally produce a graph-based camera pose trajectory along with a global 3D map of the environment composed of potentially millions of 3D points. The goal here is to produce a global 3D map which is not only composed of millions of 3D points but also 3D planes as indoor scenes are mostly composed of planar features such as floor, walls, desks and other furnitures. These planes are directly detected within the depth images provided by the sensor and therefore tend to "regularize" the noise and missing data in depth images during the whole RGB-D SLAM process. This process represents a first step towards a more semantic approach of RGB-D SLAM.

One of the key problems in Augmented Reality systems, is the synchronization of real and virtual world. In the context of our research efforts, we need to augment a view of the real world with virtual elements which first requires to determine position and orientation of the point of view and then compute a virtual view according to this location. Augmented reality uses a lot of different sensors in order to estimate camera or operator’s point of view. These sensors could provide samples faster than the mixing of virtual and real informations could be displayed. We expose here a way to take into account samples that are generated during the mixing process by combining augmented reality and virtual reality techniques. This method is using a post-rendering technique involving a texture carried out by a software graphics pipeline to perform this task mainly devoted to optical see-through augmented reality systems. This algorithm can lead to deformations on the final rendered image. We will compute the theoretical errors introduced by this method and confront these results to those obtained with a simulation test-bench which is implementing our proposal.

The major problem with augmented reality (AR) systems using see-through head mounted displays (HMD's) is the end-to-end system delay (or latency). This delay exists because the head tracker, scene generator, and communication links require time to perform their tasks, causing a lag between the measurement of head location and the display of the corresponding virtual objects inside the HMD. One way to eliminate or reduce the latency is to predict future head locations. We propose to use optimal Bayesian algorithms for non-linear/non-Gaussian tracking problems, with a focus on particle filters to predict head motion. Particle filters are sequential Monte Carlo methods based upon point mass (or 'particle') representation of probability densities, which can applied to any state space model, and which generalize the traditional Kalman filtering methods. A SIR particle filter is discussed and compared with the standard EKF through an illustrative example.

One of the key problems in augmented reality systems is registration, that is to say the synchronization of real and virtual world. Augmented reality uses a lot of different sensors in order to estimate camera or operator's point of view. These sensors could provide samples faster than mixing of virtual and real information could be displayed. We expose here a way to fake into account samples that are generated during the mixing process. This method is using a post-rendering technique involving a texture to perform this task. We expose the errors reduction obtained by performing such technique with a simulation test-bench implementing our proposal.

3D free form object recognition is one of the most difficult problems in computer vision. We present a new approach which exploits only one luminance image of a complex object to recognize it in the scene by identifying its appearance in the input image. We construct a photometric (non geometric) projective invariant to perform matching between local regions of the object in the image and those of the model. We propose an original method based on what we called "photometric aspects" to construct a discriminative data base of the 3D object model. We demonstrate the effectiveness of our approach while implementing it with complex free form objects and we present some obtained results.

Our work addresses the integration of speech recognition and natural language processing for spoken dialogue systems. To deal with recognition errors, we investigate two repairing strategies integrated in a parsing based on a Lexicalized Tree Grammar. The first strategy takes its root in the recognition hypothesis, the other in the linguistic expectations. We expose a formal framework to express the grammar, to describe the repairing strategies and to foresee further strategies.