Andre L. C. Barczak

Stream Processing of Integral Images for Real-Time Object Detection

This paper presents the design and evaluation of the stream processing implementation of the integral image algorithm. The integral image is a key component of many image processing algorithms in particular the Haar-like feature based systems. Modern GPUs provide a large number of processors with a peak floating point performance that is significantly higher than current general CPUs. This results in significant performance improvement when the Integral Image calculation for large input images is offloaded onto the GPU of the system.

Stream processing for fast and efficient rotated Haar-like features using rotated integral images

An extended set of Haar-like features for image sensors beyond the standard vertically and horizontally aligned Haar-like features and the 45° twisted Haar-like features are introduced. The extended rotated Haar-like features are based on the standard Haar-like features that have been rotated based on whole integer pixel-based rotations. These rotated feature values can also be calculated using rotated integral images which mean that they can be fast and efficiently calculated with just eight operations irrespective of the feature size. The integral image calculations can be offloaded to the graphical processing unit (GPU) using the stream processing paradigm. The integral image calculation on the GPU is seen to be faster than the traditional central processing unit implementation of the algorithm, for large image sizes, allowing more complex classifiers to be implemented in real-time.

Real-time Hand Tracking based on Non-Invariant Features

In this paper, we discuss the importance of the choice of features in digital image object recognition. The features can be classified as invariants or non-invariants. Invariant features are robust against one or more modifications such as rotations, translations, scaling and different light (illumination) conditions. Noninvariant features are usually very sensitive to any of these modifiers. On the other hand, noninvariant features can be used even in the event of translation, scaling and rotation, but the feature choice is in some cases more important than the training method. If the feature space is adequate then the training process can be straightforward and good classifiers can be obtained. In the last few years, good algorithms have been developed relying on noninvariant features. In this article, we show how noninvariant features can cope with changes even though this requires additional computation at the detection phase. We also show preliminary results for a hand detector based on a set of cooperative Haar-like feature detectors. The results show the good potential of the method as well as the challenges to achieve real-time detection

Empirical evaluation of a new structure for AdaBoost

We propose a mixed structure to form cascades for AdaBoost classifiers, where parallel strong classifiers are trained for each layer. The structure allows for rapid training and guarantees high hit rates without changing the original threshold. We implemented and tested the approach for two datasets from UCI [1], and compared results of binary classifiers using three different structures: standard AdaBoost, a cascade classifier with threshold adjustments, and the proposed structure.

Performance characteristics of a cost-effective medium-sized beowulf cluster supercomputer

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.

Classifier and Feature Based Stereo for Mobile Robot Systems

Classifier based approaches to stereo vision reduce the ambiguity associated with low level texture and feature based image registration, however there are challenges associated with providing accurate object positioning for good depth estimation using these high level approaches. This paper investigates the performance of stereo based systems that use Haar-like features for object classification. The availability of good face detectors using this approach makes it suitable for biped and mobile robot systems that operate in environments that include people, however significant challenges exist for identifying general objects that are not as highly structured and aligned as human faces.

Adaptive cascade of boosted ensembles for face detection in concept drift

We propose an adaptive learning algorithm for cascades of boosted ensembles that is designed to handle the problem of concept drift in nonstationary environments. The goal was to create a real-time adaptive algorithm for dynamic environments that exhibit varying degrees of drift in high-volume streaming data. This we achieved using a hybrid of detect-and-retrain and constant-update approaches. The uniqueness of our method is found in two aspects of our framework. The first is the manner in which individual weak classifiers within each cascade layer of an ensemble are clustered during training and assigned a competence value. Secondly, the idea of learning optimal cascade-layer thresholds during runtime, which enables rapid adaptation to dynamic environments. The proposed adaptive learning method was applied to a binary-class problem with rare-event detection characteristics. For this, we chose the domain of face detection and demonstrate experimentally the ability of our algorithm to achieve an effective trade-off between accuracy and speed of adaptations in dense data streams with unknown rates of change.

Hybrid Fuzzy Colour Processing and Learning

We present a robust fuzzy colour processing system with automatic rule extraction and colour descriptors calibration for accurate colour object recognition and tracking in real-time. The system is anchored on the fusion of fuzzy colour contrast rules that operate on the red, green and blue channels independently and adaptively to compensate for the effects of glare, shadow, and illumination variations in an indoor environment. The system also utilises a pie-slice colour classification technique in a modified rg-chromaticity space. Now, colour operations can be defined linguistically to allow a vision system to discriminate between similarly coloured objects more effectively. The validity and generality of the proposed fuzzy colour processing system is analysed by examining the complete mapping of the fuzzy colour contrast rules for each target colour object under different illumination intensities with the presence of similarly coloured objects. The colour calibration algorithm is able to extract colour descriptors in a matter of seconds as compared to manual calibration usually taking hours to complete. Using the robot soccer environment as a test bed, the algorithm is able to calibrate colours with excellent accuracy.

Stream processing of geometric and central moments using high precision summed area tables

This paper introduces a stream programming based design of the zero and higher order central moments that use an integral image or summed area data structure of geometric moments. The stream programming algorithm runs on a general purpose graphics processing unit (GPGPU) that are becoming commodity hardware, giving real-time performance even for large image sizes and a large number of scan window sizes.

Colour object classification using the fusion of visible and near-infrared spectra

Under extreme light conditions, a conventional colour CCD camera would fail to render the colours of an object properly as the visible spectrum is either faintly observable in the scene or the presence of glare corrupts the colours sensed. On the other hand, for darkly-illuminated areas, a near-infrared (NIR) camera would sense stronger more discriminable signals, but could only render the scene monochromatically. The underlying challenge in this research is how to adaptively integrate a monochromatic NIR image with a faintly rendered colour image of the same darkly or very brightly lit scene to give rise to improved colour classification results that discriminate colours more effectively. This research proposes a Fuzzy-Genetic colour processing algorithm that adaptively marries together the visible and near-infrared spectra signals for the purpose of colour object recognition. The experiments were done on a scene with spatially varying illumination intensities, using Fujifilms UV/IR Super CCD camera with a sensitivity range between 380nm to 1000nm in conjunction with NIR filters. Results prove that the proposed multi-spectrum technique yields better colour classification results than utilizing the pure visible spectrum alone.