06 Fakultät Luft- und Raumfahrttechnik und Geodäsie

Permanent URI for this collectionhttps://elib.uni-stuttgart.de/handle/11682/7

Browse

Filters

2017 - 201922020 - 20232

Settings

Institute: search.filters.UBSInstitut.Institut für PhotogrammetrieSubject: search.filters.subject.004

Search Results

Now showing 1 - 4 of 4
  • Thumbnail Image
    ItemOpen Access
    Forming a hybrid intelligence system by combining Active Learning and paid crowdsourcing for semantic 3D point cloud segmentation
    (2023) Kölle, Michael; Sörgel, Uwe (Prof. Dr.-Ing.)
    While in recent years tremendous advancements have been achieved in the development of supervised Machine Learning (ML) systems such as Convolutional Neural Networks (CNNs), still the most decisive factor for their performance is the quality of labeled training data from which the system is supposed to learn. This is why we advocate focusing more on methods to obtain such data, which we expect to be more sustainable than establishing ever new classifiers in the rapidly evolving ML field. In the geospatial domain, however, the generation process of training data for ML systems is still rather neglected in research, with typically experts ending up being occupied with such tedious labeling tasks. In our design of a system for the semantic interpretation of Airborne Laser Scanning (ALS) point clouds, we break with this convention and completely lift labeling obligations from experts. At the same time, human annotation is restricted to only those samples that actually justify manual inspection. This is accomplished by means of a hybrid intelligence system in which the machine, represented by an ML model, is actively and iteratively working together with the human component through Active Learning (AL), which acts as pointer to exactly such most decisive samples. Instead of having an expert label these samples, we propose to outsource this task to a large group of non-specialists, the crowd. But since it is rather unlikely that enough volunteers would participate in such crowdsourcing campaigns due to the tedious nature of labeling, we argue attracting workers by monetary incentives, i.e., we employ paid crowdsourcing. Relying on respective platforms, typically we have access to a vast pool of prospective workers, guaranteeing completion of jobs promptly. Thus, crowdworkers become human processing units that behave similarly to the electronic processing units of this hybrid intelligence system performing the tasks of the machine part. With respect to the latter, we do not only evaluate whether an AL-based pipeline works for the semantic segmentation of ALS point clouds, but also shed light on the question of why it works. As crucial components of our pipeline, we test and enhance different AL sampling strategies in conjunction with both a conventional feature-driven classifier as well as a data-driven CNN classification module. In this regard, we aim to select AL points in such a manner that samples are not only informative for the machine, but also feasible to be interpreted by non-experts. These theoretical formulations are verified by various experiments in which we replace the frequently assumed but highly unrealistic error-free oracle with simulated imperfect oracles we are always confronted with when working with humans. Furthermore, we find that the need for labeled data, which is already reduced through AL to a small fraction (typically ≪1 % of Passive Learning training points), can be even further minimized when we reuse information from a given source domain for the semantic enrichment of a specific target domain, i.e., we utilize AL as means for Domain Adaptation. As for the human component of our hybrid intelligence system, the special challenge we face is monetarily motivated workers with a wide variety of educational and cultural backgrounds as well as most different mindsets regarding the quality they are willing to deliver. Consequently, we are confronted with a great quality inhomogeneity in results received. Thus, when designing respective campaigns, special attention to quality control is required to be able to automatically reject submissions of low quality and to refine accepted contributions in the sense of the Wisdom of the Crowds principle. We further explore ways to support the crowd in labeling by experimenting with different data modalities (discretized point cloud vs. continuous textured 3D mesh surface), and also aim to shift the motivation from a purely extrinsic nature (i.e., payment) to a more intrinsic one, which we intend to trigger through gamification. Eventually, by casting these different concepts into the so-called CATEGORISE framework, we constitute the aspired hybrid intelligence system and employ it for the semantic enrichment of ALS point clouds of different characteristics, enabled through learning from the (paid) crowd.
  • Thumbnail Image
    ItemOpen Access
    Semi-dense filter-based visual odometry for automotive augmented reality applications
    (2019) Schmid, Stephan; Fritsch, Dieter (Prof. Dr.-Ing.)
    In order to integrate virtual objects convincingly into a real scene, Augmented Reality (AR) systems typically need to solve two problems: Firstly, the movement and position of the AR system within the environment needs to be known to be able to compensate the motion of the AR system in order to make placement of the virtual objects stable relative to the real world and to provide overall correct placement of virtual objects. Secondly, an AR system needs to have a notion of the geometry of the real environment to be able to properly integrate virtual objects into the real scene via techniques such as the determination of the occlusion relation between real and virtual objects or context-aware positioning of virtual content. To solve the second problem, the following two approaches have emerged: A simple solution is to create a map of the real scene a priori by whatever means and to then use this map in real-time operation of the AR system. A more challenging, but also more flexible solution is to create a map of the environment dynamically from real time data of sensors of the AR-system. Our target applications are Augmented Reality in-car infotainment systems in which a video of a forward facing camera is augmented. Using map data to determine the geometry of the environment of the vehicle is limited by the fact that currently available digital maps only provide a rather coarse and abstract picture of the world. Furthermore, map coverage and amount of detail vary greatly regionally and between different maps. Hence, the objective of the presented thesis is to obtain the geometry of the environment in real time from vehicle sensors. More specifically, the aim is to obtain the scene geometry by triangulating it from the camera images at different camera positions (i.e. stereo computation) while the vehicle moves. The problem of estimating geometry from camera images where the camera positions are not (exactly) known is investigated in the (overlapping) fields of visual odometry (VO) and structure from motion (SfM). Since Augmented Reality applications have tight latency requirements, it is necessary to obtain an estimate of the current scene geometry for each frame of the video stream without delay. Furthermore, Augmented Reality applications need detailed information about the scene geometry, which means dense (or semi-dense) depth estimation, that is one depth estimate per pixel. The capability of low-latency geometry estimation is currently only found in filter based VO methods, which model the depth estimates of the pixels as the state vector of a probabilistic filter (e.g. Kalman filter). However, such filters maintain a covariance matrix for the uncertainty of the pixel depth estimates whose complexity is quadratic in the number of estimated pixel depths, which causes infeasible complexity for dense depth estimation. To resolve this conflict, the (full) covariance matrix will be replaced by a matrix requiring only linear complexity in processing and storage. This way, filter-based VO methods can be combined with dense estimation techniques and efficiently scaled up to arbitrarily large image sizes while allowing easy parallelization. For treating the covariance matrix of the filter state, two methods are introduced and discussed. These methods are implemented as modifications to the (existing) VO method LSD-SLAM, yielding the "continuous" variant C-LSD-SLAM. In the first method, a diagonal matrix is used as the covariance matrix. In particular, the correlation between different scene point estimates is neglected. For stabilizing the resulting VO method in forward motion, a reweighting scheme is introduced based on how far scene point estimates are moved when reprojecting them from one frame to the next frame. This way, erroneous scene point estimates are prevented from causing the VO method to diverge. The second method for treating the covariance matrix models the correlation of the scene point estimates caused by camera pose uncertainty by approximating the combined influence of all camera pose estimates in a small subspace of the scene point estimates. This subspace has fixed dimension 15, which forces the complexity of the replacement of the covariance matrix to be linear in the number of scene point estimates.
  • Thumbnail Image
    ItemOpen Access
    Concept and performance evaluation of a novel UAV-borne topo-bathymetric LiDAR sensor
    (2020) Mandlburger, Gottfried; Pfennigbauer, Martin; Schwarz, Roland; Flöry, Sebastian; Nussbaumer, Lukas
    We present the sensor concept and first performance and accuracy assessment results of a novel lightweight topo-bathymetric laser scanner designed for integration on Unmanned Aerial Vehicles (UAVs), light aircraft, and helicopters. The instrument is particularly well suited for capturing river bathymetry in high spatial resolution as a consequence of (i) the low nominal flying altitude of 50-150 m above ground level resulting in a laser footprint diameter on the ground of typically 10-30 cm and (ii) the high pulse repetition rate of up to 200 kHz yielding a point density on the ground of approximately 20-50 points/m2. The instrument features online waveform processing and additionally stores the full waveform within the entire range gate for waveform analysis in post-processing. The sensor was tested in a real-world environment by acquiring data from two freshwater ponds and a 500 m section of the pre-Alpine Pielach River (Lower Austria). The captured underwater points featured a maximum penetration of two times the Secchi depth. On dry land, the 3D point clouds exhibited (i) a measurement noise in the range of 1-3 mm; (ii) a fitting precision of redundantly captured flight strips of 1 cm; and (iii) an absolute accuracy of 2-3 cm compared to terrestrially surveyed checkerboard targets. A comparison of the refraction corrected LiDAR point cloud with independent underwater checkpoints exhibited a maximum deviation of 7.8 cm and revealed a systematic depth-dependent error when using a refraction coefficient of n = 1.36 for time-of-flight correction. The bias is attributed to multi-path effects in the turbid water column (Secchi depth: 1.1 m) caused by forward scattering of the laser signal at suspended particles. Due to the high spatial resolution, good depth performance, and accuracy, the sensor shows a high potential for applications in hydrology, fluvial morphology, and hydraulic engineering, including flood simulation, sediment transport modeling, and habitat mapping.
  • Thumbnail Image
    ItemOpen Access
    Development of a SGM-based multi-view reconstruction framework for aerial imagery
    (2017) Rothermel, Mathias; Fritsch, Dieter (Prof. Dr.-Ing.)
    Advances in the technology of digital airborne camera systems allow for the observation of surfaces with sampling rates in the range of a few centimeters. In combination with novel matching approaches, which estimate depth information for virtually every pixel, surface reconstructions of impressive density and precision can be generated. Therefore, image based surface generation meanwhile is a serious alternative to LiDAR based data collection for many applications. Surface models serve as primary base for geographic products as for example map creation, production of true-ortho photos or visualization purposes within the framework of virtual globes. The goal of the presented theses is the development of a framework for the fully automatic generation of 3D surface models based on aerial images - both standard nadir as well as oblique views. This comprises several challenges. On the one hand dimensions of aerial imagery is considerable and the extend of the areas to be reconstructed can encompass whole countries. Beside scalability of methods this also requires decent processing times and efficient handling of the given hardware resources. Moreover, beside high precision requirements, a high degree of automation has to be guaranteed to limit manual interaction as much as possible. Due to the advantages of scalability, a stereo method is utilized in the presented thesis. The approach for dense stereo is based on an adapted version of the semi global matching (SGM) algorithm. Following a hierarchical approach corresponding image regions and meaningful disparity search ranges are identified. It will be verified that, dependent on undulations of the scene, time and memory demands can be reduced significantly, by up to 90% within some of the conducted tests. This enables the processing of aerial datasets on standard desktop machines in reasonable times even for large fields of depth. Stereo approaches generate disparity or depth maps, in which redundant depth information is available. To exploit this redundancy, a method for the refinement of stereo correspondences is proposed. Thereby redundant observations across stereo models are identified, checked for geometric consistency and their reprojection error is minimized. This way outliers are removed and precision of depth estimates is improved. In order to generate consistent surfaces, two algorithms for depth map fusion were developed. The first fusion strategy aims for the generation of 2.5D height models, also known as digital surface models (DSM). The proposed method improves existing methods regarding quality in areas of depth discontinuities, for example at roof edges. Utilizing benchmarks designed for the evaluation of image based DSM generation we show that the developed approaches favorably compare to state-of-the-art algorithms and that height precisions of few GSDs can be achieved. Furthermore, methods for the derivation of meshes based on DSM data are discussed. The fusion of depth maps for 3D scenes, as e.g. frequently required during evaluation of high resolution oblique aerial images in complex urban environments, demands for a different approach since scenes can in general not be represented as height fields. Moreover, depths across depth maps possess varying precision and sampling rates due to variances in image scale, errors in orientation and other effects. Within this thesis a median-based fusion methodology is proposed. By using geometry-adaptive triangulation of depth maps depth-wise normals are extracted and, along the point coordinates are filtered and fused using tree structures. The output of this method are oriented points which then can be used to generate meshes. Precision and density of the method will be evaluated using established multi-view benchmarks. Beside the capability to process close range datasets, results for large oblique airborne data sets will be presented. The report closes with a summary, discussion of limitations and perspectives regarding improvements and enhancements. The implemented algorithms are core elements of the commercial software package SURE, which is freely available for scientific purposes.