The bundle of geometry and appearance in computer vision has proven to be a promising solution for robots across a wide variety of applications. Stereo cameras and RGB-D sensors are widely used to realise fast 3D reconstruction and trajectory tracking in a dense way. However, they lack flexibility of seamless switch between different scaled environments, i.e., indoor and outdoor scenes. In addition, semantic information are still hard to acquire in a 3D mapping. We address this challenge by combining the stateof-art deep learning method and semi-dense Simultaneous Localisation and Mapping (SLAM) based on video stream from a monocular camera. In our approach, 2D semantic information are transferred to 3D mapping via correspondence between connective Keyframes with spatial consistency. There is no need to obtain a semantic segmentation for each frame in a sequence, so that it could achieve a reasonable computation time. We evaluate our method on indoor/outdoor datasets and lead to an improvement in the 2D semantic labelling over baseline single frame predictions.

No abstract

No abstract

Accurate localization of a vehicle is a challenging task as GPS available on the market are not designed for lane-level accuracy application. Although dead reckoning helps, cumulative errors from inertial sensors result in a integration drift. This paper presents a new method of localization based on sensors data fusion. An accurate digital map of the lane marking is used as a powerful additional sensor. Road markings are detected by processing two lateral cameras to estimate their distance to the vehicle. Coupled with the map data in a EKF filter it improves the ego-localization obtained with inertial and GPS measurements. The result is a vehicle localization at an ego-lane level of accuracy, with a lateral error of less than 10 centimeters. I. IDriver Assistance Systems developed over the last decade have required a precise and robust estimation of road scene major features. Those features include obstacles (vehicle, pedestrian), road (marking, lanes, traffic signs), and the egovehicle (localization and dynamics of the vehicle). Usually each feature is addressed separately, for instance obstacle detection in collision avoidance, road attributes in lane keeping assistance, ego-localization in navigation systems. Recently, automated driving systems have made it necessary to fuse the attributes of different features to obtain more precise, robust and complete information.In the frame of french, european and international projects (respectively ABV, eFuture, CooPerCom) we have tackled the task of perception of the environment including the attributes of these three features. More specifically, we develop an application localizing the ego-vehicle in its lane allowing a positioning and a lateral control precise enough to be applicable.There is a body of work in the field of robust localization by hybrid data fusion (proprio and exteroceptive): monomodel approaches (EKF, UKF, DD1, DD2) [1], [2], [3], multi-model [4], [5] and particular filter [6] achieve localization with a precision to the meter. For instance, in [5], the approach is only centered on the ego-vehicle but can compute the likelihood of each model to build the finale estimation. In [6] a map is used to filter out the particles out of a road. In this case, the localization process converges rapidly toward a solution inside a roadway. However it cannot specify which lane the vehicle is on.It is clear that the use of a geographic map can be an advantage: commercial navigation systems routinely operate MapMatching by coupling maps and vehicle positioning. In [7] a map-matching algorithm using the visual odometry motion trajectory estimation (from a stereovision rig) as input Authors are with IFSTTAR, COSYS-LIVIC, Fig. 1.Extrinsic configuration of the side cameras and associated coordinate systems. and corrected using the digital map features, greatly improve the global localization performance. In [8], a standard navigation map is matched with a laser scanner occupancy grid map, a video based grid map and a lane marking grid map for road course pr...

Road sign identification in images is an important issue, in particular for vehicle safety applications. It is usually tackled in three stages: detection, recognition and tracking, and evaluated as a whole. To progress towards better algorithms, we focus in this paper on the first stage of the process, namely road sign detection. More specifically, we compare, on the same ground-truth image database, results obtained by three algorithms that sample different state-of-the-art approaches. The three tested algorithms: Contour Fitting, Radial Symmetry Transform, and pair-wise voting scheme, all use color and edge information and are based on geometrical models of road signs. The test dataset is made of 847 images 960 × 1080 of complex urban scenes (available at www.itowns.fr/benchmarking.html). They feature 251 road signs of different shapes (circular, rectangular, triangular), sizes and types. The pros and cons of the three algorithms are discussed, allowing to draw new research perspectives.