region-based human object recognition using angular radial transform

Abstract

Human object recognition is an active research topic in visual surveillance. This thesis develops a robust method for region-based human recognition with angular radial transform (ART). The main idea of this thesis is computing the Euclidean distance between the different ART vectors to judge the object is human or not. This approach is based on huge images dataset. The boundary smooth and mean-shift are applied to improve the recognition rate. In the training dataset, 2000 extracted human-body-shape images which are achieved by the background subtraction (BS) and 200 illumination human-body-shape images are applied to learn with ART. All these 2200 images in the training dataset will be normalized to the size of 32×32, the noise will be reduced and the boundary of the human body shape will be smoothed. ART is a complex orthogonal unitary transform defined on a unit disk that consists of the complete orthogonal sinu-soidal basis functions in polar coordinates [5], ART coefficients are computed by projecting the input image onto the ART basis functions. The similarity of two shapes can be determined by the Euclidean distance between two ART vectors. In this thesis, the object is human or not is determined by comparing the ART vector of input image and the ART vector of threshold-the value of threshold is decided by the , and it can be adjusted by the parameter α. In query, 1000 different shapes of images are chosen, including 500 extracted human, 300 illumination images of MPEG CE-2-B database, 100 illumination animals, 50 extracted cares and 50 illumination plants. Based on this test dataset, the recognition rate is 73.5%, the false negative is 82.8% and the false positive is 64.6%. In order to improve the recognition rate, the mean-shift based clustering is applied to classify the dataset. In the training set, the 2200 images are classified into 6 clusters. By testing, the optimum α could be achieved for different clusters. After clustering, the recognition rate raises from 73.5% to 85.1%. The result will show that the proposed algorithm is robust, flexible and efficient.