Supervised Local High-Order Differential Channel Feature Learning for Pedestrian Detection

In this paper, a novel supervised local high-order differential channel feature is proposed for fast pedestrian detection. This method is motivated by the recent successful use of filtering on the multiple channel maps, which can improve the performance. This method firstly compute the multiple channel maps for the input RGB image, and average pooling is acted on the channel maps in order to reduce the effect of noise and sample misalignment. Then, each of the pooled channel maps is convolved with our proposed local high-order filter bank, which can enhance the discriminative information in the feature space. Finally, due to the increasing memory consumption incurred by the higher dimension of resulting feature, we have proposed a local structure preserved supervised dimension reduction method which aims to keep the manifold structure of samples in the feature space. This method is formulated as a classical spectral graph embedding problem which can be solved by the LPP algorithms. Thorough experiments and comparative studies show that our method can achieve very competitive result compared with many state-of-art methods on the INRIA and Caltech datasets. Besides, our detector can run about 20 fps in 480 \(\times \) 640 resolution images.