Michiel Vlaminck

Obstacle detection for pedestrians with a visual impairment based on 3D imaging

According to the World Health Organisation, 285 million people live with a visual impairment. Despite the fact that many efforts have been made recently, there is still no computer-guided system that is reliable, robust and practical enough to help these people to increase their mobility. Motivated by this shortcoming, we propose a novel obstacle detection system to assist the visually impaired. This work mainly focuses on indoor environments and performs classification of typical obstacles that emerge in these situations, using a 3D sensor. A total of four classes of obstacles are considered: walls, doors, stairs and a residual class (which covers loose obstacles and bumpy parts on the floor). The proposed system is very reliable in terms of the detection accuracy. In a realistic experiment, stairs are detected with 100% true positive rate and 8.6% false positive rate, while doors are detected with 86.4% true positive rate and 0% false positive rate.

Towards online mobile mapping using inhomogeneous lidar data

In this paper we present a novel approach to quickly obtain detailed 3D reconstructions of large scale environments. The method is based on the consecutive registration of 3D point clouds generated by modern lidar scanners such as the Velodyne HDL-32e or HDL-64e. The main contribution of this work is that the proposed system specifically deals with the problem of sparsity and inhomogeneity of the point clouds typically produced by these scanners. More specifically, we combine the simplicity of the traditional iterative closest point (ICP) algorithm with the analysis of the underlying surface of each point in a local neighbourhood. The algorithm was evaluated on our own collected dataset captured with accurate ground truth. The experiments demonstrate that the system is producing highly detailed 3D maps at the speed of 10 sensor frames per second.

Indoor assistance for visually impaired people using a RGB-D camera

In this paper a navigational aid for visually impaired people is presented. The system uses a RGB-D camera to perceive the environment and implements self-localization, obstacle detection and obstacle classification. The novelty of this work is threefold. First, self-localization is performed by means of a novel camera tracking approach that uses both depth and color information. Second, to provide the user with semantic information, obstacles are classified as walls, doors, steps and a residual class that covers isolated objects and bumpy parts on the floor. Third, in order to guarantee real time performance, the system is accelerated by offloading parallel operations to the GPU. Experiments demonstrate that the whole system is running at 9 Hz.

3D Scene Reconstruction Using Omnidirectional Vision and LiDAR: A Hybrid Approach.

In this paper, we propose a novel approach to obtain accurate 3D reconstructions of large-scale environments by means of a mobile acquisition platform. The system incorporates a Velodyne LiDAR scanner, as well as a Point Grey Ladybug panoramic camera system. It was designed with genericity in mind, and hence, it does not make any assumption about the scene or about the sensor set-up. The main novelty of this work is that the proposed LiDAR mapping approach deals explicitly with the inhomogeneous density of point clouds produced by LiDAR scanners. To this end, we keep track of a global 3D map of the environment, which is continuously improved and refined by means of a surface reconstruction technique. Moreover, we perform surface analysis on consecutive generated point clouds in order to assure a perfect alignment with the global 3D map. In order to cope with drift, the system incorporates loop closure by determining the pose error and propagating it back in the pose graph. Our algorithm was exhaustively tested on data captured at a conference building, a university campus and an industrial site of a chemical company. Experiments demonstrate that it is capable of generating highly accurate 3D maps in very challenging environments. We can state that the average distance of corresponding point pairs between the ground truth and estimated point cloud approximates one centimeter for an area covering approximately 4000 m2. To prove the genericity of the system, it was tested on the well-known Kitti vision benchmark. The results show that our approach competes with state of the art methods without making any additional assumptions.

3D Object Finding Using Geometrical Constraints on Depth Images

Finding an object in a 3D scene is an important problem in the robotics, especially in assistive systems for visually impaired people. In most systems, the first and most important step is how to detect an object in a complex environment. In this paper, we propose a method for finding an object using geometrical constraints on depth images from a Kinect. The main advantage of the approach is it is invariant to lighting condition, color and texture of the objects. Our approach does not require a training phase, therefore it can reduce the time of preparing data and learning model. The objects of interest have a simple geometrical structure such as coffee mugs, bowls, boxes and are on a table. Overall, our approach is faster and more accurate than methods using 2D features on depth images for training an object model.

Multi-resolution ICP for the efficient registration of point clouds based on octrees

In this paper we propose a multiresolution scheme based on hierarchical octrees for the registration of point clouds acquired by lidar scanners. The point density of these point clouds is generally sparse and inhomogeneous, a property that can yield a risk for correct alignment. Experiments demonstrate that our multiresolution technique is a lot faster than the traditional iterative closest point (ICP) algorithm while it is more robust, e.g. in case of abrupt movements of the sensor. We can report a speed-up factor of more than 30, without jeopardizing the level of accuracy. In scenarios for which the level of detail is less critical, e.g. in case of navigation for autonomous robots, we can even achieve a larger speed-up by trading speed for quality.

Liborg: a lidar-based robot for efficient 3D mapping

In this work we present Liborg, a spatial mapping and localization system that is able to acquire 3D models on the y using data originated from lidar sensors. The novelty of this work is in the highly efficient way we deal with the tremendous amount of data to guarantee fast execution times while preserving sufficiently high accuracy. The proposed solution is based on a multi-resolution technique based on octrees. The paper discusses and evaluates the main benefits of our approach including its efficiency regarding building and updating the map and its compactness regarding compressing the map. In addition, the paper presents a working prototype consisting of a robot equipped with a Velodyne Lidar Puck (VLP-16) and controlled by a Raspberry Pi serving as an independent acquisition platform.

Consistent ICP for the registration of sparse and inhomogeneous point clouds

In this paper, we derive a novel iterative closest point (ICP) technique that performs point cloud alignment in a robust and consistent way. Traditional ICP techniques minimize the point-to-point distances, which are successful when point clouds contain no noise or clutter and moreover are dense and more or less uniformly sampled. In the other case, it is better to employ point-to-plane or other metrics to locally approximate the surface of the objects. However, the point-to-plane metric does not yield a symmetric solution, i.e. the estimated transformation of point cloud p to point cloud q is not necessarily equal to the inverse transformation of point cloud q to point cloud p. In order to improve ICP, we will enforce such symmetry constraints as prior knowledge and make it also robust to noise and clutter. Experimental results show that our method is indeed much more consistent and accurate in presence of noise and clutter compared to existing ICP algorithms.