<> "The repository administrator has not yet configured an RDF license."^^ . <> . . "Compound Models for Vision-Based Pedestrian Recognition"^^ . "This thesis addresses the problem of recognizing pedestrians in video images acquired from a moving camera in real-world cluttered environments. Instead of focusing on the development of novel feature primitives or pattern classifiers, we follow an orthogonal direction and develop feature- and classifier-independent compound techniques which integrate complementary information from multiple image-based sources with the objective of improved pedestrian classification performance. After establishing a performance baseline in terms of a thorough experimental study on monocular pedestrian recognition, we investigate the use of multiple cues on module-level. A motion-based focus of attention stage is proposed based on a learned probabilistic pedestrian-specific model of motion features. The model is used to generate pedestrian localization hypotheses for subsequent shape- and texture-based classification modules. In the remainder of this work, we focus on the integration of complementary information directly into the pattern classification step. We present a combination of shape and texture information by means of pose-specific generative shape and texture models. The generative models are integrated with discriminative classification models by utilizing synthesized virtual pedestrian training samples from the former to enhance the classification performance of the latter. Both models are linked using Active Learning to guide the training process towards informative samples. A multi-level mixture-of-experts classification framework is proposed which involves local pose-specific expert classifiers operating on multiple image modalities and features. In terms of image modalities, we consider gray-level intensity, depth cues derived from dense stereo vision and motion cues arising from dense optical flow. We furthermore employ shape-based, gradient-based and texture-based features. The mixture-of-experts formulation compares favorably to joint space approaches, in view of performance and practical feasibility. Finally, we extend this mixture-of-experts framework in terms of multi-cue partial occlusion handling and the estimation of pedestrian body orientation. Our occlusion model involves examining occlusion boundaries which manifest in discontinuities in depth and motion space. Occlusion-dependent weights which relate to the visibility of certain body parts focus the decision on unoccluded body components. We further apply the pose-specific nature of our mixture-of-experts framework towards estimating the density of pedestrian body orientation from single images, again integrating shape and texture information. Throughout this work, particular emphasis is laid on thorough performance evaluation both regarding methodology and competitive real-world datasets. Several datasets used in this thesis are made publicly available for benchmarking purposes. Our results indicate significant performance boosts over state-of-the-art for all aspects considered in this thesis, i.e. pedestrian recognition, partial occlusion handling and body orientation estimation. The pedestrian recognition performance in particular is considerably advanced; false detections at constant detection rates are reduced by significantly more than an order of magnitude."^^ . "2011" . . . . . . . . "Markus"^^ . "Enzweiler"^^ . "Markus Enzweiler"^^ . . . . . . "Compound Models for Vision-Based Pedestrian Recognition (PDF)"^^ . . . "enzweiler_phd_print.pdf"^^ . . . "Compound Models for Vision-Based Pedestrian Recognition (Other)"^^ . . . . . . "indexcodes.txt"^^ . . . "Compound Models for Vision-Based Pedestrian Recognition (Other)"^^ . . . . . . "lightbox.jpg"^^ . . . "Compound Models for Vision-Based Pedestrian Recognition (Other)"^^ . . . . . . "preview.jpg"^^ . . . "Compound Models for Vision-Based Pedestrian Recognition (Other)"^^ . . . . . . "medium.jpg"^^ . . . "Compound Models for Vision-Based Pedestrian Recognition (Other)"^^ . . . . . . "small.jpg"^^ . . "HTML Summary of #12099 \n\nCompound Models for Vision-Based Pedestrian Recognition\n\n" . "text/html" . . . "510 Mathematik"@de . "510 Mathematics"@en . .