Pedestrian detection with an ensemble of localized features

In this paper, a new human detection approach from still image is proposed. Two vector features are extracted from the image. Histogram of oriented gradient feature represents the gradient information of human. Histogram of modified local binary pattern is extracted from images convolved with Gabor filter, as a feature vector to represent texture information. It can be seen as a supplement of gradient information. Different support vector machine classifiers are trained by each type of vectors. Finally, two classifiers are combined together for the final result by using the proposed integration method. Because two features contain different information, they have low error dependency and can get high detection rate. Experiment is performed in a large dataset and it shows that this method outperforms state-of-the-art approaches and other combinations of features.