Abstract: By using the constraints around the neighborhoods for weighted sparse representation， the pedestrian detection problem was solved. A dictionary with a strong extracting discriminate and sparse features power was obtained by using the Fisher discriminant dictionary learning method. With the constraint of the neighborhoods， the image patch was represented as a sparse feature via the dictionary. By computing the representation of the residuals and comparing the residuals with a threshold， the patch label was determined to finish the classification task. The experiments on INRIA person datasets showed that non-local sparse feature has an obvious power of discrimination. The constraint of the neighborhoods makes the sparse feature represented effectively.

Key words: pedestrian detection, nonlocal, sparse representation, discriminate dictionary；optimization ore matching