Dense scene reconstruction for autonomous robot navigation

Details

Supervisors

Ronsse, Renaud ; Haelterman, Rob

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électromécanicien, à finalité spécialisée: mécatronique

Abstract

enThis master thesis addresses the problem of 3D scene reconstruction for autonomous robots that navigate in adversarial environments. In such environments, stealth is valuable, which motivates using cameras as the main sensor instead of alternatives such as Light Detection and Ranging (LIDAR) sensors, that easily give up their location by emitting signals. Cameras deliver sequences of images, and efficient algorithms can already extract, for each image in the sequence, the 3D pose of the camera. Additionally, those algorithms extract a sparse point cloud as a representation of the environment. However, this sparse representation of the environment does not provide enough information for tasks such as autonomous navigation. This observation motivates studying 3D reconstruction algorithms that provide richer representation of the environment, and use as input the sequence of image from the camera whose poses are provided by existing algorithms. Navigation algorithms need an up-to-date representation of the environment. To achieve this, a 3D reconstruction algorithm working in real time is necessary. The first step in this direction is to build a working algorithm, that can then be optimized for real-time operation. Therefore, the goal of this master thesis is to build 3D reconstruction algorithms that can ultimately be adapted for real-time. The environment is represented by assigning a depth to each pixel of one image, which is selected from the sequence of images given by the camera. This depth assignment is called the depth map, and it is found as the minimum of an energy function. The energy function is the sum of a data term and a regularization term. The data term promotes depth maps that are in accordance with the data from the posed image sequence. The regularization term promotes depth maps that exhibit structures expected in the true depth map. This master thesis analyzes the state-of-the-art regularization terms, while using only a simple data term for all the methods studied. Experiments are performed on various environments from a photo-realistic simulated dataset, which provides the ground truth inverse depth maps. Two novel regularization methods are introduced: one based on second-order-derivatives of the inverse depth map, and one based on wavelet transforms. They both show competitive reconstruction results on the simulated dataset compared to the re-implemented state-of-the-art methods. Additionally, the wavelet-based method is promising for future optimization due to the simple iterative solver designed to solve it, and its potential for parallelization. This work is amenable to many extensions. In particular, to investigate whether more advanced data terms can improve the presented methods or not. Another valuable extension would be to test the proposed methods closer to real conditions: with images from a real camera and poses given as the output of an existing algorithm.