Object Detection by Learning Oriented Gradients

Jiajie Chen, Huicheng Zheng, Na He, Ziquan Luo, Rui Zhu

Abstract: This paper proposes a method of learning features corresponding to oriented gradients for efficient object detection. Instead of dividing a local patch into cells with fixed sizes and locations such as in the traditional HOG, we employ a data-driven method to learn the sizes and locations of cells. Firstly, oriented gradient patch-maps of a local patch are constructed according to the orientations. Secondly, rectangular cells of various sizes and locations are constructed in each patch-map to sum up the magnitudes of oriented gradients and produce candidate local features. The local features are then selected by using a boosting procedure. Finally, a local patch is represented by a feature vector in which each component corresponds to the sum of oriented gradients in a rectangular cell. An object detector is then trained over the local patches by using a higher-level boosted cascade structure. Extensive experimental results on public datasets verified the superiority of the proposed method to existing related methods in terms of both the training speed and the detection accuracy.