Eugen Funk

Recursive Total Variation Filtering Based 3D Fusion

3D reconstruction from mobile image sensors is crucial for many offline-inspection and online robotic application. While several techniques are known today to deliver high accuracy 3D models from images via offline-processing, 3D reconstruction in real-time remains a major goal still to achieve. This work focuses on incremental 3D modeling from error prone depth image data, since standard 3D fusion techniques are tailored on accurate depth data from active sensors such as the Kinect. Imprecise depth data is usually provided by stereo camera sensors or simultaneous localization and mapping (SLAM) techniques. This work proposes an incremental extension of the total variation (TV) filtering technique, which is shown to reduce the errors of the reconstructed 3D model by up to 77% compared to state of the art techniques.

Infinite, Sparse 3D Modelling Volumes

Modern research in mobile robotics proposes to combine localization and perception in order to recognize previously visited locations and thus to improve localization as well as the object recognition processes recursively. A crucial issue is to perform updates of the scene geometry when novel observations become available. The reason is that a practical application often requires a system to model large 3D environments at high resolution which exceeds the storage of the local memory. The underlying work presents an optimized volume data structure for infinite 3D environments which facilitates (i) successive world model updates without the need to recompute the full dataset, (ii) very fast in-memory data access scheme enabling the integration of high resolution 3D sensors in real-time, (iii) efficient level-of-detail for visualization and coarse geometry updates. The technique is finally demonstrated on real world application scenarios which underpin the feasibility of the research outcomes.

IPS – a vision aided navigation system

Abstract Ego localization is an important prerequisite for several scientific, commercial, and statutory tasks. Only by knowing one’s own position, can guidance be provided, inspections be executed, and autonomous vehicles be operated. Localization becomes challenging if satellite-based navigation systems are not available, or data quality is not sufficient. To overcome this problem, a team of the German Aerospace Center (DLR) developed a multi-sensor system based on the human head and its navigation sensors – the eyes and the vestibular system. This system is called integrated positioning system (IPS) and contains a stereo camera and an inertial measurement unit for determining an ego pose in six degrees of freedom in a local coordinate system. IPS is able to operate in real time and can be applied for indoor and outdoor scenarios without any external reference or prior knowledge. In this paper, the system and its key hardware and software components are introduced. The main issues during the development of such complex multi-sensor measurement systems are identified and discussed, and the performance of this technology is demonstrated. The developer team started from scratch and transfers this technology into a commercial product right now. The paper finishes with an outlook.

Boundless Reconstruction Using Regularized 3D Fusion

3D reconstruction from image based depth sensor is essential part of many offline or online robotic applications. Numerous techniques have been developed to integrate multiple depth maps to create 3D model of environment, however accuracy of the reconstructed 3D model exclusively depends upon the precision of depth sensing. Economical depth sensors such as Kinect and stereo camera sensors provide imprecise depth data which affect the integration process and produce unwanted noisy surfaces in 3D model. There exist several approaches which use image filtering based depth map denoising, however applying filtering directly on depth data can result in inconsistent and deformed 3D model. In this paper we investigate and extend a recursive variant of total variation based filtering to incorporate multi-view based depth images while applying implicit depth smoothing. Proposed framework uses sparse voxel representation to aid large scale 3D model reconstruction and is shown to reduce absolute surface error of final reconstructed 3D model by up to 77% in comparison with state of the art 3D fusion techniques.

Infinite 3D Modelling Volumes

Modern research in mobile robotics proposes to combine localization and perception in order to recognize previously visited locations and thus to improve localization as well as the object recognition processes recursively. A crucial issue is to perform updates of the scene geometry when novel observations become available. The reason is that a practical application often requires a system to model large 3D environments at high resolution which exceeds the storage of the local memory. The underlying work presents an optimized volume data structure for infinite 3D environments which facilitates i) successive world model updates without the need to recompute the full dataset, ii) very fast in-memory data access scheme enabling the integration of high resolution 3D sensors in real-time, iii) efficient level-of-detail for visualization and coarse geometry updates. The technique is finally demonstrated on real world application scenarios which underpin the feasibility of the research outcomes.

TVL1 Shape Approximation from Scattered 3D Data

With the emergence in 3D sensors such as laser scanners and 3D reconstruction from cameras, large 3D point clouds can now be sampled from physical objects within a scene. The raw 3D samples delivered by these sensors however, contain only a limite d degree of information about the environment the objects exist in, which means that further geometrical high-level modelling is essential. In addition, issues like sparse data measurements, noise, missing samples due to occlusion, and the inherently huge datasets involved in such representations makes this task extremely challenging. This paper addresses these issues by presenting a new 3D shape modelling framework for samples acquired from 3D sensor. Motivated by the success of nonlinear kernel-based approximation techniques in the statistics domain, existing methods using radial basis functions are applied to 3D object shape approximation. The task is framed as an optimization problem and is extended using non-smooth L1 total variation regularization. Appropriate convex energy functionals are constructed and solved by applying the Alternating Direction Method of Multipliers approach, which is then extended using Gauss-Seidel iterations. This significantly lowers the computational complexity involved in generating 3D shape from 3D samples, while both numerical and qualitative analysis confirms the superior shape modelling performance of this new framework compared with existing 3D shape reconstruction techniques.

TVL{}_1 Planarity Regularization for 3D Shape Approximation

The modern emergence of automation in many industries has given impetus to extensive research into mobile robotics. Novel perception technologies now enable cars to drive autonomously, tractors to till a field automatically and underwater robots to construct pipelines. An essential requirement to facilitate both perception and autonomous navigation is the analysis of the 3D environment using sensors like laser scanners or stereo cameras. 3D sensors generate a very large number of 3D data points when sampling object shapes within an environment, but crucially do not provide any intrinsic information about the environment which the robots operate within.