Refine
Document Type
- Conference Proceeding (9) (remove)
Language
- English (9)
Has Fulltext
- no (9)
Is part of the Bibliography
- no (9)
In this paper, we describe a method to model human clothes for a later recognition by the use of RGB- and SWIR-cameras. A basic model is estimated during people detection and tracking. This model will be refined if the recognition is triggered. For the refining, several saliency maps are used to extract individual features. These individual features are located separately for any human body parts. The body parts are estimated by the use of a silhouette extraction combined with a skeleton estimation. In this way, the model describes the human clothes in a compact manner which allows the use of a simple and fast comparison method for people recognition. Such models can be used in security and service applications.
In this paper, we describe an efficient method for a fast people re-identification based on models of human clothes. An initial model is estimated during people detection and tracking, which will be refined during the re-identification. This stepwise extraction, combination and comparing of features speeds up the whole re-identification. For the refining, several saliency maps are used to extract individual features. These individual features are located separately for any human body part. The body parts are located with an optimized GPU-based HOG detector. Furthermore, we introduce a meanshift-based fusion concept which utilizes multiple detectors in order to increase the detection reliability.
We present a novel approach of distributing matrix multiplications among GPU-equipped nodes in a cluster system. In this context we discuss the induced challenges and possible solutions. Additionally we state an algorithm which outperforms optimized GPU BLAS libraries for small matrices. Furthermore we provide a novel theoretical model for distributing algorithms within homogeneous computation systems with multiple hierarchies. In the context of this model we develop an algorithm which can find the optimal distribution parameters for each involved subalgorithm. We provide a detailed analysis of the algorithms space and time complexities and justify its use with a structured evaluation within a small GPU-equipped Beowulf cluster.
In the context of existing approaches to cluster computing we present a newly developed modular framework `SimpleHydra' for rapid deployment and management of Beowulf clusters. Instead of focusing only the pure computation tasks on homogeneous clusters (i.e. clusters with identically set up nodes), this framework aims to ease the configuration of heterogeneous clusters and to provide a low-level / high-level object-oriented API for low-latency distributed computing. Our framework does not make any restrictions regarding the hardware and minimizes the use of external libraries to the case of special modules. In addition to that our framework enables the user to develop highly dynamic cluster topologies. We describe the framework's general structure as well as time critical elements, give application examples in the `Big-Data' context during a research project and briefly discuss additional features. Furthermore we give a thorough theoretical time/space complexity analysis of our implemented methods and general approaches.
Practical application of object detection systems, in research or industry, favors highly optimized black box solutions. We show how such a highly optimized system can be further augmented in terms of its reliability with only a minimal increase of computation times, i.e. preserving realtime boundaries. Our solution leaves the initial (HOG-based) detector unchanged and introduces novel concepts of non-linear metrics and fusion of ROIs. In this context we also introduce a novel way of combining feature vectors for mean-shift grouping. We evaluate our approach on a standarized image database with a HOG detector, which is representative for practical applications. Our results show that the amount of false-positive detections can be reduced by a factor of 4 with a negligable complexity increase. Although introduced and applied to a HOG-based system, our approach can easily be adapted for different detectors.
Object detection systems which operate on large data streams require an efficient scaling with available computation power. We analyze how the use of tile-images can increase the efficiency (i.e. execution speed) of distributed HOG-based object detectors. Furthermore we discuss the challenges of using our developed algorithms in practical large scale scenarios. We show with a structured evaluation that our approach can provide a speed-up of 30-180 % for existing architectures. Due to the its generic formulation it can be applied to a wide range of HOG-based (or similar) algorithms. In this context we also study the effects of applying our method to an existing detector and discuss a scalable strategy for distributing the computation among nodes in a cluster system.
We present a novel approach of distributing small-to mid-scale neural networks onto modern parallel architectures. In this context we discuss the induced challenges and possible solutions. We provide a detailed theoretical analysis with respect to space and time complexities and reinforce our computation model with evaluations which show a performance gain over state of the art approaches.