Section: New Results

3D Scene Mapping

New RGB-D sensor design for indoor 3D mapping

Participants : Eduardo Fernandez Moral, Patrick Rives.

A multi-sensor device has been developed for omnidirectional RGB-D (color+depth) image acquisition (see Fig. 3 .a). This device allows to acquire such omnidirectional images at high frame rates (30 Hz). This approach has advantages over other alternatives used today in terms of accuracy and real-time spherical image construction for indoor environments, which are specially interesting for mobile robotics. This device has important prospective applications as fast 3D-reconstruction or Slam.

A calibration method for such device was developed [31] , which takes into account the bias of each sensor independently. The proposed calibration method does not require any specific calibration pattern, taking into account the planar structure from the scene to cope with the fact that there is no overlapping between sensors.

In a first instance, this sensor has been exploited for localization and mapping research with mobile robots. For that, the sensor is mounted on a mobile platform together with an standard computer (see Fig. 3 .a). A method to perform image registration and visual odometry has been developed. This method relies in the matching of planar primitives that can be efficiently obtained from the depth images. This technique performs considerably faster than previous registration approaches like ICP, or dense photoconsistency alignment. These last achieve however a better accuracy than our method, what suggests that our method can be used as an initial step to speed-up those.

Slam is also addressed with this device. A solution to this problem using our omnidirectional RGB-D sensor is being researched. The ongoing experiments have shown some initial results for metric-topological pose-graph Slam, where the map consists of a set of spherical keyframes, which are located in a topological arrangement according to their shared observations.

Compact 3D scene representation

Participants : Renato José Martins, Patrick Rives, Tawsif Gokhool.

This work follows in the direction of precise and compact scene representation of large scale environments. The aim is to build a complete geometric and photometric “minimal” model, which is stored within a sparse set of augmented spherical images to asset photo-geometry consistence of the scene from multiple points-of-views. In this direction, an uncertainty model from the full structure combined with those of poses was proposed for point-to-point egocentric fusion. This model allows to reduce sensor noise in a particular keyframe sphere when performing a multi-frame fusion scheme of coherent near information. This first fusion scheme is then improved by exploiting the rigidity/influence of neighboring points representing the surface. For that, an intermediary higher level abstraction of the point cloud is generated by partitioning the input domain into elementary cells, then reducing the number of degrees of freedom and enforcing constraints over the points segmented as being part of the same surface.

The adopted solution is a “weaker” representation of a 3D boundary mesh, based on discontinuous convex planar patches, with the segmentation being done considering the geometry (region growing) or photometry (SLIC superpixels). This synthetic scene built with the planar geometric police proved to well represent the original scene (for both indoor and outdoor real data) with a significant small amount of patches and it is exploited to build robust useful “dynamic” 4D world model, which in turn can be used for assisted/autonomous navigation or virtual reality applications.

Semantic mapping

Participants : Romain Drouilly, Patrick Rives, Panagiotis Papadakis.

Autonomous navigation is one of the most challenging problems to address to allow robots to evolve in our everyday environments. Map-based navigation has been studied for a long time and researches have produced a great variety of approaches to model the world. However, semantic information has only recently been taken into account in those models to improve robot efficiency [56] . The goal of this work is to study how semantics can be used to improve all the steps of navigation process. In a first time, we have developed a new navigation-oriented hybrid metric-topological-semantic model of the world. It captures high-level information and uses it to build extremely compact description of large environments. Then we have used it to design an efficient localization algorithm, able to find a given map content faster than classical methods and allowing human-understandable queries [30] . In a second time, we have studied how semantics can be used to discover unobserved things in the scene. Particularly, we have shown that both statics and dynamic entities, identified by a robot, can inform about the structure of the environment in unobserved areas [29] . We have used this to do “map extrapolation”, that is extending a map beyond robot's perceptual limits by reasoning on semantics. This approach has been shown to be of great interest in everyday-life environment. Finally, we have proposed a new scheme for trajectory planing, taking into account not only geometric constraints but also high-level understanding of the world. We have shown the usefulness of this approach to navigate complex environments with highly dynamic areas on both simulated and real-world datasets, well-suited for large outdoor environment navigation.

Augmented reality

Participant : Eric Marchand.

Using Slam methods becomes more and more common in Augmented Reality (AR). To achieve real-time requirement and to cope with scale factor and the lack of absolute positioning issue, we proposed to decouple the localization and the mapping step. This approach has been validated on an Android Smartphone through a collaboration with Orange Labs [38] [39]