Fast Pedestrian Detection Based on the Selective Window Differential Filter

• Published in: Neural processing letters Vol. 48; no. 1; pp. 403 - 417
• Main Authors: Zuo, Xin, Shen, Jifeng, Yu, Hualong, Xu, Dan, Qian, Chengshan, Shan, Yongwei
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2018; Springer Nature B.V
• Subjects: Accuracy; Artificial Intelligence; Complex Systems; Computational Intelligence; Computer Science; Deep learning; Feature selection; Filtration; Pixels; Random variables; Reduction; Feature selection; Windows differential feature; Pedestrian detection
• ISSN: 1370-4621; 1573-773X
• DOI: 10.1007/s11063-017-9746-8

Following the recent progress of the pixel-level filtering for pedestrian detection, we propose a window differential feature (WDF) based on the multiple channel maps. More specifically, WDF encodes first-order statistics between artitary two pixels in the whole detection window, thus obtaining larger receptive field for achor pixel than other filtering methods. Despite obtaining more discriminative information for pedestrian, WDF suffers expensive space complexity due to the high feature dimensionality. Quantitive analysis for the arbitrary pairwise elements in the WDF vector demonstrates the weak correlations existing in the proposed feature, thus motivate dimension reduction with feature selection to be the top choice. Three different dimension reduction methods for the WDF demonstrate that feature selection with mutual information achieves superior result. In addition, we find the complementary characteristics between the baseline feature and selective window differential feature, thus combining both can obtain further performance improvement. Extensive experiments on the INRIA, Caltech, ETH, and TUD-Brussel datasets consistently show superior performance of the proposed method to state-of-the-art methods with a 22 fps running speed for 640 × 480 images.