Pedestrian detection in traffic scenes using multi-attitude classifiers

We propose a novel algorithm that detects pedestrians based on their body appearance. As a pedestrian has a high variance in shape we create a star based classification scheme that contains a cascaded root classifier (trained on multiple attitudes) and four classifiers trained on specific pedestrian attitudes (rear, front, lateral left and lateral right). We use Histogram of Oriented Gradient features and Local Binary Patterns that are extracted on parts positioned along different regions of the pedestrian model. The parts are composed of several blocks that are chosen such that a homogeneity function of edge variation is minimized. The block based approach is useful for capturing the variations in shape and position of pedestrian body parts. The novelty of our method resides in the combination of multi-attitude classification model with the usage of block-based feature extraction.