Exploring Prior Knowledge for Pedestrian Detection

Pedestrian detection is a classical and hot issue in object detection. Many approaches have been proposed in this area. However, it remains a challenging problem due to the variances in lighting conditions, scene structures, clothes, view angles, postures, scales, occlusions, etc. Previous survey [1] has summarized that using better features plays an important role in improving detection quality. In addition, prior knowledge has shown good success in designing haar-like features for pedestrian detection [4]. Inspired by it, our work aims to integrate more prior knowledge into the design of features to enhance performance of pedestrian detection. By observing the pedestrian samples, we have discovered several important priors that are always ignored by previous methods, e.g. the symmetry of human body and the differences among different channels. We therefore utilize these priors to design two kinds of features: 1) symmetric features which capture the difference between two local symmetric regions, and 2) cross-channel features which capture the difference between two different channels of the same region. Figure 1 gives some visual examples of these two features. To the best of our knowledge, we are the first to use symmetric and cross-channel priors in designing features for pedestrian detection. By integrating these prior information into feature design, our detector achieves state-of-the-art performance.