Computer Vision News - February 2019

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen Visual SLAM for Automated Driving: Exploring the Applications of Deep Learning . We are indebted to the authors ( Stefan Milz, Georg Arbeiter, Christian Witt, Bassam Abdallah and Senthil Yogamani ), for allowing us to use their images. It is the Valeo's Visual SLAM pipeline for Autonomous Driving. Credit for the algorithm goes to the Valeo Kronach team. The paper is here . Deep Learning has become the go-to solution for tasks like detection and recognition, and lately there have been advances in using CNNs for geometric tasks, depth assessment, optical flow prediction and motion segmentation. However, Visual SLAM (Self-Localization and Mapping) for Automated Driving systems is a field in which CNN methods are not yet sufficiently developed for production. In this article, the authors review the field and outline where and how Deep Learning can likely be applied to replace some of the stages of Visual SLAM. The paper start by describing the challenges, building-blocks and stages of Visual SLAM. Next, the applications of Visual SLAM to Automated Driving is reviewed. And finally, opportunities for applying Deep Learning to improve upon the classic methods currently in use is discussed. The challenges of Visual SLAM Algorithm challenges: ● Depth assessment: If the on-car camera rotates, depth-assessment from consecutive images is almost impossible. ● World Map and initialization: Most approaches to world-mapping start from a random initialization -- which creates many challenges as the rate of convergence is dependent on camera motion, ● Scale Ambiguity: A single camera SLAM can only estimate the World Map and Trajectory up to a certain scale, and requires a global reference to get a precise scale. ● Rolling shutter: This image acquisition method scans the frame horizontally or vertically, which results in not all parts of the input image being acquired simultaneously. This can create distortions. (as in the “bent” propeller blade in the image) 4 Research - Visual SLAM for Automated Driving Research by Assaf Spanier Computer Vision News CNNs are quickly becoming the go-to solution for vision tasks like object detection and semantic segmentation for Automated Driving