A new technique for iconic hypothesis verification in model-based vision systems has been developed, which enhances the resolution of the problem of three-dimensional object recognition in two-dimensional scenes. This thesis investigates an iconic feature-matching approach to verification, in which two-dimensional image features are predicted from a specific view of a three-dimensional geometric model, and these features are matched directly to the unprocessed image data. This solves the crucial image to model registration problem. The iconic matching approach solves two of the major disadvantages of the usual symbolic matching method; where symbolic image constructs are compared with symbolic model data. The symbolic description of image features is not robust, and detailed matches cannot be made, as much of the original data has been lost. The investigation of iconic verification has been split into two parts. Firstly individual features are matched. Secondly the results from these are aggregated into a model match score. For the first stage four iconic evaluators have been developed and compared. These predictive evaluators are designed to assess the "edge-ness" of a small patch of an image. The advantage of one of these techniques over its equivalent data-driven approach is shown. The complete verification procedure aggregates the image-specific iconic feature evaluation scores. The iconic matching technique has been tested in the domain of car recognition in outdoor scene images. Its sensitivity in images containing a great deal of distracting noise has been very encouraging. There are however many application areas for this research. Iconic matching can be used to track both individual features and entire objects, for example in successive frames of a sequence of images over time

No description available.

No description available.

Many tasks require surveillance and analysis in order to make decisions regarding the next course of action. The people responsible for these tasks are usually concerned with any event that affects their bottom-line. Traditionally, human operators have had to either actively man a set of video displays to determine if specific events were occurring or manually review hours of collected video data to see if a specific event occurred. Actively monitoring video stream or manually reviewing and analyzing the data collected, however, is a tedious and long process which is prone to errors due to biases and inattention. Automatically processing and analyzing the video provides an alternate way of getting more accurate results because it can reduce the likelihood of missing important events and the human factors that lead to decreased efficiency. The thesis aims to contribute to the area of using computer vision as a decision support tool by integrating detector, tracker, re-identification, activity status estimation, and event processor modules to generate the necessary event statistics needed by a human operator. The contribution of this thesis is a system that uses feedback from each of the modules to provide better target detection, and tracking results for event statistics generation over an extended period of time. To demonstrate the efficacy of the proposed system, it is first used to generate event statistics that measure productivity on multiple construction work sites. The versatility of the proposed system is also demonstrated in an indoor assisted living environment by using it to determine how much of an influence a technology intervention had on promoting interactions amongst older adults in a shared space.

published_or_final_version / Electrical and Electronic Engineering / Doctoral / Doctor of Philosophy

published_or_final_version / Electrical and Electronic Engineering / Master / Master of Philosophy

published_or_final_version / Psychology / Master / Master of Philosophy

This thesis explains the development of a computer vision system able to find and orient relatively small objects. The motivations is exchanging a monotonous work done by hand and replace it with an automation system with help of an ABB IRB 140 industrial robot. The vision system runs on a standard PC and is developed using the OpenCV environment, originally made by Intel in Russia. The algorithms of the system is written in C++ and the user interface in C++/CLI. With a derived test case, multiple vision algorithms is tested and evaluated for this kind of application. The result shows that SIFT/SURF works poorly with multiple instances of the search object and HAAR classifiers produces many false positives. Template matching with image moment calculation gave a satisfying result regarding multiple object in the scene and produces no false positives. Drawbacks of the selected algorithm developed where sensibility to light invariance and lack of performance in a skewed scene. The report also contains suggestions on how to precede with further improvements or research.

Although stereoscopic depth perception has been a common subject of research since 1838, it is only recently that a fundamental division between two types of stereoscopic processing -- first-order and second-order -- has been identified. The precise functioning of the neural mechanisms that underlie these processing streams, however, is not yet understood. It was first confirmed that first-order stereopsis mechanisms are best investigated with vertically oriented stimuli at small disparities , whereas second-order stereopsis mechanisms prefer stimuli at large disparities. It has been postulated that the second-order stereopsis mechanisms are insensitive to stimulus spatial frequency content, but this was found not to be the case. Four different explanations were investigated to account for this result: nonlinearities of the early visual system and screen, absolute and relative orientation of the stimuli, and the possibility of the second-order mechanism being multiple. The last of these possibilities was found to be the most plausible, based on the data collected. It has long been thought that vergence may play an important role in stereopsis because both systems use disparity cues. This relationship was investigated using briefly presented "second-order" stimuli and both experienced and naive subjects. A disparity sign judgement (front/back) to disparate stimuli was made while vergence eye movements were recorded. Stimuli were presented over a range of orientations, presentation times, disparities and contrasts. The vergence responses were found to be uncorrelated with the disparity sign judgements and it was concluded that vergence does not influence stereopsis under these conditions. Finally, the role that first- and second-order stereopsis mechanisms play within tilt percepts induced by orientation disparities were investigated. Subjects were presented with two different types of stimuli with either carrier, envelope, or both carrier and envelope orientation disparities. The stimulus with no "first-order reference" could not be perceived in depth with the envelope orientation disparity alone. When this stimulus had both carrier and envelope orientation disparities, depth tilt was still perceived. It was concluded that the perception of stereoscopic tilt is a first-order process.

No description available.