Illumination invariant interest point detection for vision based recognition tasks

Vision based recognition systems learn the appearance of given objects using images. These objects can be recognised and localised in other images after camera motion and illumination changes. The goal of this work is to improve the ability of such systems to recognise objects after illumination changes. Recognition systems usually reduce the amount of image data used for recognition by detecting interest points: small characteristic image patches. Most interest point detectors are sensitive to illumination changes. Illumination invariant interest point detection would increase the proportion of points which are redetected after illumination changes and it would decrease the proportion of false points. It would hence improve the performance of recognition systems. Several new interest point detectors with higher stability under illumination changes are developed in this work. They are based on the Harris detector, which is often used because of its stability under viewpoint changes. Four new detectors are developed for grey value images. They all adapt detection to the local lighting intensity using different principles: local normalisation, homomorphic processing, local threshold adaptation and local clustering. Two new detectors are developed for colour images. The first one adapts detection to local lighting intensity and colour with homomorphic processing. The second detector is based on an invariant colour space. It fully eliminates light intensity influence and it locally compensates light colour. In addition, an appropriate demosaicing method and a preprocessing based on the Nagao filter are presented, in order to reduce the influence of noise and colour artifacts on the colour detectors. Detection stability is evaluated for all new detectors and for the existing Harris detector on image series acquired under varying illumination. The new detectors are more stable than the Harris detector for scenes with complex 3D geometry, for non–uniform lighting and for complex illumination changes such as light source movement. The best results are obtained with the homomorphic detectors and, if the scene has good colour edges, with the detector based on the invariant colour space. A robust state of the art recognition system is also developed and used to evaluate the detectors in a practical application. Systems using colour information perform better than systems based on grey values. For grey value images, the new homomorphic detector improves recognition performances for complex objects or it increases the system efficiency, depending on the chosen thresholding method. For colour systems, the new detectors achieve higher performance improvements than for grey value systems. The detector based on the invariant colour space achieves the best recognition performances if the object contains good colour edges. The homomorphic colour detector also performs very well and is suitable for all kinds of objects. The developed algorithms improve hence detection stability and recognition performances.