An improved Haar-like feature for efficient object detection

Abstract In this paper, we propose an improved feature descriptor, Haar Contrast Feature , for efficient object detection under various illumination conditions. The proposed feature uses the same prototypes of Haar-like feature and computes contrast using the normalization factor devised to reflect the average intensity of feature region. It is computed efficiently using an integral image and is more powerful in real-time applications by not requiring variance normalization during detection process. It shows improved performance under a wide range of illumination conditions. For experiments, classifiers for face, pedestrian, and vehicle were trained by employing the conventional Haar-like feature with/without variance normalization, the local binary pattern descriptor, and the proposed feature descriptor, and their performances were evaluated. Experimental results confirm that classifiers employing the proposed feature descriptor outperform those employing the conventional Haar-like feature or the local binary pattern descriptor in terms of detection accuracy under most illumination conditions.