Browsing by Author "Barczak, A.L.C."

Now showing 1 - 7 of 7
  • Thumbnail Image
    Item
    Accelerated face detector training using the PSL framework
    (Massey University, 2009) Susnjak, T.; Barczak, A.L.C.; Hawick, K.A.
    We train a face detection system using the PSL framework [1] which combines the AdaBoost learning algorithm and Haar-like features. We demonstrate the ability of this framework to overcome some of the challenges inherent in training classifiers that are structured in cascades of boosted ensembles (CoBE). The PSL classifiers are compared to the Viola-Jones type cas- caded classifiers. We establish the ability of the PSL framework to produce classifiers in a complex domain in significantly reduced time frame. They also comprise of fewer boosted en- sembles albeit at a price of increased false detection rates on our test dataset. We also report on results from a more diverse number of experiments carried out on the PSL framework in order to shed more insight into the effects of variations in its adjustable training parameters.
  • Thumbnail Image
    Item
    Face tracking using a hyperbolic catadioptric omnidirectional system
    (Massey University, 2009) Barczak, A.L.C.; Okamoto Jr, J.; Grassi Jr, V.
    In the first part of this paper, we present a brief review on catadioptric omnidirectional systems. The special case of the hyperbolic omnidirectional system is analysed in depth. The literature shows that a hyperboloidal mirror has two clear advantages over alternative geometries. Firstly, a hyperboloidal mirror has a single projection centre [1]. Secondly, the image resolution is uniformly distributed along the mirror’s radius [2]. In the second part of this paper we show empirical results for the detection and tracking of faces from the omnidirectional images using Viola-Jones method. Both panoramic and perspective projections, extracted from the omnidirectional image, were used for that purpose. The omnidirectional image size was 480x480 pixels, in greyscale. The tracking method used regions of interest (ROIs) set as the result of the detections of faces from a panoramic projection of the image. In order to avoid losing or duplicating detections, the panoramic projection was extended horizontally. Duplications were eliminated based on the ROIs established by previous detections. After a confirmed detection, faces were tracked from perspective projections (which are called virtual cameras), each one associated with a particular face. The zoom, pan and tilt of each virtual camera was determined by the ROIs previously computed on the panoramic image. The results show that, when using a careful combination of the two projections, good frame rates can be achieved in the task of tracking faces reliably.
  • Thumbnail Image
    Item
    A new 2D static hand gesture colour image dataset for ASL gestures
    (Massey University, 2011) Barczak, A.L.C.; Reyes, N.H.; Abastillas, M.; Piccio, A.; Susnjak, T.
    It usually takes a fusion of image processing and machine learning algorithms in order to build a fully-functioning computer vision system for hand gesture recognition. Fortunately, the complexity of developing such a system could be alleviated by treating the system as a collection of multiple sub-systems working together, in such a way that they can be dealt with in isolation. Machine learning need to feed on thousands of exemplars (e.g. images, features) to automatically establish some recognisable patterns for all possible classes (e.g. hand gestures) that applies to the problem domain. A good number of exemplars helps, but it is also important to note that the efficacy of these exemplars depends on the variability of illumination conditions, hand postures, angles of rotation, scaling and on the number of volunteers from whom the hand gesture images were taken. These exemplars are usually subjected to image processing first, to reduce the presence of noise and extract the important features from the images. These features serve as inputs to the machine learning system. Different sub-systems are integrated together to form a complete computer vision system for gesture recognition. The main contribution of this work is on the production of the exemplars. We discuss how a dataset of standard American Sign Language (ASL) hand gestures containing 2425 images from 5 individuals, with variations in lighting conditions and hand postures is generated with the aid of image processing techniques. A minor contribution is given in the form of a specific feature extraction method called moment invariants, for which the computation method and the values are furnished with the dataset.
  • Thumbnail Image
    Item
    A novel bootstrapping method for positive datasets in cascades of boosted ensembles
    (Massey University, 2010) Susnjak, T.; Barczak, A.L.C.; Hawick, K.A.
    We present a novel method for efficiently training a face detector using large positive datasets in a cascade of boosted ensembles. We extend the successful Viola-Jones [1] framework which achieved low false acceptance rates through bootstrapping negative samples with the capability to also bootstrap large positive datasets thereby capturing more in-class variation of the target object. We achieve this form of bootstrapping by way of an additional embedded cascade within each layer and term the new structure as the Bootstrapped Dual-Cascaded (BDC) framework. We demonstrate its ability to easily and efficiently train a classifier on large and complex face datasets which exhibit acute in-class variation.
  • Thumbnail Image
    Item
    Performance characteristics of a cost-effective medium-sized Beowulf cluster supercomputer
    (2003) Barczak, A.L.C.; Messom, C.H.; Johnson, M.J.
    This paper presents some performance results obtained from a new Beowulf cluster, the Helix, built at Massey University, Auckland funded by the Allan Wilson Center for Evolutionary Ecology. Issues concerning network latency and the effect of the switching fabric and network topology on performance are discussed. In order to assess how the system performed using the message passing interface (MPI), two test suites (mpptest and jumpshot) were used to provide a comprehensive network performance analysis. The performance of an older fast-ethernet/single processor based cluster is compared to the new Gigabit/SMP cluster. The Linpack performance of Helix is investigated. The Linpack Rmax rating of 234.8 Gflops puts the cluster at third place in the Australia/ New Zealand sublist of the Top500 supercomputers, an extremely good performance considering the commodity parts and its low cost (US$125000).
  • Thumbnail Image
    Item
    Real-time computation of Haar-like features at generic angles for detection algorithms
    (Massey University, 2006) Barczak, A.L.C.; Johnson, M.J.; Messom, C.H.
    This paper proposes a new approach to detect rotated objects at distinct angles using the Viola-Jones detector. The use of additional Integral Images makes an approximation the Haar-like features for any given angle. The proposed approach uses di erent types of Haar-like features, including features that compute areas at 45o, 26.5o and 63.5o of rotation. Given a trained classi er (using normal features) a conversion is made using a pair of features so an equivalent value is computed for any angle. This conversion is only an approximation, but the errors are constrained and they would have limited impact on the nal accuracy of the classi er. We discuss the sources of errors in the computation of the Haar-like features and show through experiments that in natural images the errors are often negligible.
  • Thumbnail Image
    Item
    A reconfigurable hybrid intelligent system for robot navigation
    (Massey University, 2011) Reyes, N.H.; Barczak, A.L.C.; Fatahillah; Susnjak, T.
    Soft computing has come of age to o er us a wide array of powerful and e cient algorithms that independently matured and in uenced our approach to solving problems in robotics, search and optimisation. The steady progress of technology, however, induced a ux of new real-world applications that demand for more robust and adaptive computational paradigms, tailored speci cally for the problem domain. This gave rise to hybrid intelligent systems, and to name a few of the successful ones, we have the integration of fuzzy logic, genetic algorithms and neural networks. As noted in the literature, they are signi cantly more powerful than individual algorithms, and therefore have been the subject of research activities in the past decades. There are problems, however, that have not succumbed to traditional hybridisation approaches, pushing the limits of current intelligent systems design, questioning their solutions of a guarantee of optimality, real-time execution and self-calibration. This work presents an improved hybrid solution to the problem of integrated dynamic target pursuit and obstacle avoidance, comprising of a cascade of fuzzy logic systems, genetic algorithm, the A* search algorithm and the Voronoi diagram generation algorithm.