| This thesis introduces a novel local appearance method, termed coherency filtering, which allows for the robust detection and analysis of rigid objects contained in heterogeneous scenes by properly exploiting the wealth of information returned by a k-nearest neighbours (k-NN) classifier. A significant advantage of k-NN classifiers is their ability to indicate uncertainty in the classification of a local window by returning a list of k candidate classifications. Classification of a local window can be inherently uncertain when considered in isolation since local windows from different objects or the background may be similar in appearance. In order to robustly identify objects in a query image, a process is needed to appropriately resolve this uncertainty. Coherency filtering resolves this uncertainty by imposing constraints across the colour channels of a query image along with spatial constraints between neighbouring local windows in a manner that produces reliable classification of local windows and ultimately results in the robust identification of objects.; Extensive experimental results demonstrate that the proposed system can robustly identify objects contained in test images focusing on pose, scale, illumination, occlusion, and image noise. A qualitative comparison with four state-of-the-art systems indicates comparable or superior performance on test sets of similar difficulty can be achieved by the proposed system, while being capable of robustly identifying objects under a greater range of viewing conditions. |
