Wide-baseline keypoint detection and matching with wide-angle images for vision based localisation

This thesis addresses the problem of detecting and describing the same scene points in different wide-angle images taken by the same camera at different viewpoints. This is a core competency of many vision-based localisation tasks including visual odometry and visual place recognition. Wide-angle cameras have a large field of view that can exceed a full hemisphere, and the images they produce contain severe radial distortion. When compared to traditional narrow field of view perspective cameras, more accurate estimates of camera egomotion can be found using the images obtained with wide-angle cameras. The ability to accurately estimate camera egomotion is a fundamental primitive of visual odometry, and this is one of the reasons for the increased popularity in the use of wide-angle cameras for this task. Their large field of view also enables them to capture images of the same regions in a scene taken at very different viewpoints, and this makes them suited for visual place recognition. However, the ability to estimate the camera egomotion and recognise the same scene in two different images is dependent on the ability to reliably detect and describe the same scene points, or ‘keypoints’, in the images. Most algorithms used for this purpose are designed almost exclusively for perspective images. Applying algorithms designed for perspective images directly to wide-angle images is problematic as no account is made for the image distortion. The primary contribution of this thesis is the development of two novel keypoint detectors, and a method of keypoint description, designed for wide-angle images. Both reformulate the Scale- Invariant Feature Transform (SIFT) as an image processing operation on the sphere. As the image captured by any central projection wide-angle camera can be mapped to the sphere, applying these variants to an image on the sphere enables keypoints to be detected in a manner that is invariant to image distortion. Each of the variants is required to find the scale-space representation of an image on the sphere, and they differ in the approaches they used to do this. Extensive experiments using real and synthetically generated wide-angle images are used to validate the two new keypoint detectors and the method of keypoint description. The best of these two new keypoint detectors is applied to vision based localisation tasks including visual odometry and visual place recognition using outdoor wide-angle image sequences. As part of this work, the effect of keypoint coordinate selection on the accuracy of egomotion estimates using the Direct Linear Transform (DLT) is investigated, and a simple weighting scheme is proposed which attempts to account for the uncertainty of keypoint positions during detection. A word reliability metric is also developed for use within a visual ‘bag of words’ approach to place recognition.

Ontology-based image annotation

With regard to the long-standing problem of the semantic gap between low-level image features and high-level human knowledge, the image retrieval community has recently shifted its emphasis from low-level features analysis to high-level image semantics extrac- tion. User studies reveal that users tend to seek information using high-level semantics. Therefore, image semantics extraction is of great importance to content-based image retrieval because it allows the users to freely express what images they want. Semantic content annotation is the basis for semantic content retrieval. The aim of image anno- tation is to automatically obtain keywords that can be used to represent the content of images. The major research challenges in image semantic annotation are: what is the basic unit of semantic representation? how can the semantic unit be linked to high-level image knowledge? how can the contextual information be stored and utilized for image annotation? In this thesis, the Semantic Web technology (i.e. ontology) is introduced to the image semantic annotation problem. Semantic Web, the next generation web, aims at mak- ing the content of whatever type of media not only understandable to humans but also to machines. Due to the large amounts of multimedia data prevalent on the Web, re- searchers and industries are beginning to pay more attention to the Multimedia Semantic Web. The Semantic Web technology provides a new opportunity for multimedia-based applications, but the research in this area is still in its infancy. Whether ontology can be used to improve image annotation and how to best use ontology in semantic repre- sentation and extraction is still a worth-while investigation. This thesis deals with the problem of image semantic annotation using ontology and machine learning techniques in four phases as below. 1) Salient object extraction. A salient object servers as the basic unit in image semantic extraction as it captures the common visual property of the objects. Image segmen- tation is often used as the �rst step for detecting salient objects, but most segmenta- tion algorithms often fail to generate meaningful regions due to over-segmentation and under-segmentation. We develop a new salient object detection algorithm by combining multiple homogeneity criteria in a region merging framework. 2) Ontology construction. Since real-world objects tend to exist in a context within their environment, contextual information has been increasingly used for improving object recognition. In the ontology construction phase, visual-contextual ontologies are built from a large set of fully segmented and annotated images. The ontologies are composed of several types of concepts (i.e. mid-level and high-level concepts), and domain contextual knowledge. The visual-contextual ontologies stand as a user-friendly interface between low-level features and high-level concepts. 3) Image objects annotation. In this phase, each object is labelled with a mid-level concept in ontologies. First, a set of candidate labels are obtained by training Support Vectors Machines with features extracted from salient objects. After that, contextual knowledge contained in ontologies is used to obtain the �nal labels by removing the ambiguity concepts. 4) Scene semantic annotation. The scene semantic extraction phase is to get the scene type by using both mid-level concepts and domain contextual knowledge in ontologies. Domain contextual knowledge is used to create scene con�guration that describes which objects co-exist with which scene type more frequently. The scene con�guration is represented in a probabilistic graph model, and probabilistic inference is employed to calculate the scene type given an annotated image. To evaluate the proposed methods, a series of experiments have been conducted in a large set of fully annotated outdoor scene images. These include a subset of the Corel database, a subset of the LabelMe dataset, the evaluation dataset of localized semantics in images, the spatial context evaluation dataset, and the segmented and annotated IAPR TC-12 benchmark.

Person-independent facial expression detection using constrained local models

In automatic facial expression detection, very accurate registration is desired which can be achieved via a deformable model approach where a dense mesh of 60-70 points on the face is used, such as an active appearance model (AAM). However, for applications where manually labeling frames is prohibitive, AAMs do not work well as they do not generalize well to unseen subjects. As such, a more coarse approach is taken for person-independent facial expression detection, where just a couple of key features (such as face and eyes) are tracked using a Viola-Jones type approach. The tracked image is normally post-processed to encode for shift and illumination invariance using a linear bank of filters. Recently, it was shown that this preprocessing step is of no benefit when close to ideal registration has been obtained. In this paper, we present a system based on the Constrained Local Model (CLM) which is a generic or person-independent face alignment algorithm which gains high accuracy. We show these results against the LBP feature extraction on the CK+ and GEMEP datasets.

User-driven saliency maps for evaluating Region-of-Interest detection

Detection of Region of Interest (ROI) in a video leads to more efficient utilization of bandwidth. This is because any ROIs in a given frame can be encoded in higher quality than the rest of that frame, with little or no degradation of quality from the perception of the viewers. Consequently, it is not necessary to uniformly encode the whole video in high quality. One approach to determine ROIs is to use saliency detectors to locate salient regions. This paper proposes a methodology for obtaining ground truth saliency maps to measure the effectiveness of ROI detection by considering the role of user experience during the labelling process of such maps. User perceptions can be captured and incorporated into the definition of salience in a particular video, taking advantage of human visual recall within a given context. Experiments with two state-of-the-art saliency detectors validate the effectiveness of this approach to validating visual saliency in video. This paper will provide the relevant datasets associated with the experiments.

Textures of optical flow for real-time anomaly detection in crowds

Automated visual surveillance of crowds is a rapidly growing area of research. In this paper we focus on motion representation for the purpose of abnormality detection in crowded scenes. We propose a novel visual representation called textures of optical flow. The proposed representation measures the uniformity of a flow field in order to detect anomalous objects such as bicycles, vehicles and skateboarders; and can be combined with spatial information to detect other forms of abnormality. We demonstrate that the proposed approach outperforms state-of-the-art anomaly detection algorithms on a large, publicly-available dataset.

Comparison of standard image segmentation methods for segmentation of brain tumors from 2D MR images

In the analysis of medical images for computer-aided diagnosis and therapy, segmentation is often required as a preliminary step. Medical image segmentation is a complex and challenging task due to the complex nature of the images. The brain has a particularly complicated structure and its precise segmentation is very important for detecting tumors, edema, and necrotic tissues in order to prescribe appropriate therapy. Magnetic Resonance Imaging is an important diagnostic imaging technique utilized for early detection of abnormal changes in tissues and organs. It possesses good contrast resolution for different tissues and is, thus, preferred over Computerized Tomography for brain study. Therefore, the majority of research in medical image segmentation concerns MR images. As the core juncture of this research a set of MR images have been segmented using standard image segmentation techniques to isolate a brain tumor from the other regions of the brain. Subsequently the resultant images from the different segmentation techniques were compared with each other and analyzed by professional radiologists to find the segmentation technique which is the most accurate. Experimental results show that the Otsu’s thresholding method is the most suitable image segmentation method to segment a brain tumor from a Magnetic Resonance Image.

Particulate texture image analysis with applications

Texture analysis and textural cues have been applied for image classification, segmentation and pattern recognition. Dominant texture descriptors include directionality, coarseness, line-likeness etc. In this dissertation a class of textures known as particulate textures are defined, which are predominantly coarse or blob-like. The set of features that characterise particulate textures are different from those that characterise classical textures. These features are micro-texture, macro-texture, size, shape and compaction. Classical texture analysis techniques do not adequately capture particulate texture features. This gap is identified and new methods for analysing particulate textures are proposed. The levels of complexity in particulate textures are also presented ranging from the simplest images where blob-like particles are easily isolated from their back- ground to the more complex images where the particles and the background are not easily separable or the particles are occluded. Simple particulate images can be analysed for particle shapes and sizes. Complex particulate texture images, on the other hand, often permit only the estimation of particle dimensions. Real life applications of particulate textures are reviewed, including applications to sedimentology, granulometry and road surface texture analysis. A new framework for computation of particulate shape is proposed. A granulometric approach for particle size estimation based on edge detection is developed which can be adapted to the gray level of the images by varying its parameters. This study binds visual texture analysis and road surface macrotexture in a theoretical framework, thus making it possible to apply monocular imaging techniques to road surface texture analysis. Results from the application of the developed algorithm to road surface macro-texture, are compared with results based on Fourier spectra, the auto- correlation function and wavelet decomposition, indicating the superior performance of the proposed technique. The influence of image acquisition conditions such as illumination and camera angle on the results was systematically analysed. Experimental data was collected from over 5km of road in Brisbane and the estimated coarseness along the road was compared with laser profilometer measurements. Coefficient of determination R2 exceeding 0.9 was obtained when correlating the proposed imaging technique with the state of the art Sensor Measured Texture Depth (SMTD) obtained using laser profilometers.

Dynamic texture reconstruction from sparse codes for unusual event detection in crowded scenes

Unusual event detection in crowded scenes remains challenging because of the diversity of events and noise. In this paper, we present a novel approach for unusual event detection via sparse reconstruction of dynamic textures over an overcomplete basis set, with the dynamic texture described by local binary patterns from three orthogonal planes (LBPTOP). The overcomplete basis set is learnt from the training data where only the normal items observed. In the detection process, given a new observation, we compute the sparse coefficients using the Dantzig Selector algorithm which was proposed in the literature of compressed sensing. Then the reconstruction errors are computed, based on which we detect the abnormal items. Our application can be used to detect both local and global abnormal events. We evaluate our algorithm on UCSD Abnormality Datasets for local anomaly detection, which is shown to outperform current state-of-the-art approaches, and we also get promising results for rapid escape detection using the PETS2009 dataset.

Unusual Event Detection in Crowded Scenes Using Bag of LBPs in spatio-temporal patches

Modelling events in densely crowded environments remains challenging, due to the diversity of events and the noise in the scene. We propose a novel approach for anomalous event detection in crowded scenes using dynamic textures described by the Local Binary Patterns from Three Orthogonal Planes (LBP-TOP) descriptor. The scene is divided into spatio-temporal patches where LBP-TOP based dynamic textures are extracted. We apply hierarchical Bayesian models to detect the patches containing unusual events. Our method is an unsupervised approach, and it does not rely on object tracking or background subtraction. We show that our approach outperforms existing state of the art algorithms for anomalous event detection in UCSD dataset.

Visual voice activity detection using frontal versus profile views

Visual activity detection of lip movements can be used to overcome the poor performance of voice activity detection based solely in the audio domain, particularly in noisy acoustic conditions. However, most of the research conducted in visual voice activity detection (VVAD) has neglected addressing variabilities in the visual domain such as viewpoint variation. In this paper we investigate the effectiveness of the visual information from the speaker’s frontal and profile views (i.e left and right side views) for the task of VVAD. As far as we are aware, our work constitutes the first real attempt to study this problem. We describe our visual front end approach and the Gaussian mixture model (GMM) based VVAD framework, and report the experimental results using the freely available CUAVE database. The experimental results show that VVAD is indeed possible from profile views and we give a quantitative comparison of VVAD based on frontal and profile views The results presented are useful in the development of multi-modal Human Machine Interaction (HMI) using a single camera, where the speaker’s face may not always be frontal.

Detection versus false alarm characterisation of a vision-based airborne dim-target collision detection system

This paper presents a preliminary flight test based detection range versus false alarm performance characterisation of a morphological-hidden Markov model filtering approach to vision-based airborne dim-target collision detection. On the basis of compelling in-flight collision scenario data, we calculate system operating characteristic (SOC) curves that concisely illustrate the detection range versus false alarm rate performance design trade-offs. These preliminary SOC curves provide a more complete dim-target detection performance description than previous studies (due to the experimental difficulties involved, previous studies have been limited to very short flight data sample sets and hence have not been able to quantify false alarm behaviour). The preliminary investigation here is based on data collected from 4 controlled collision encounters and supporting non-target flight data. This study suggests head-on detection ranges of approximately 2.22 km under blue sky background conditions (1.26 km in cluttered background conditions), whilst experiencing false alarms at a rate less than 1.7 false alarms/hour (ie. less than once every 36 minutes). Further data collection is currently in progress.