Specific Patches Decorrelation

Channel Feature on Pedestrian Detection

Xue-ming Ding and Dong-fei Ji

School of Optical Electrical and Computer Engineering, University of Shanghai for Science and Technology, China

Abstract: Typical Local Decorrelation Channel Feature (LDCF) for pedestrian detection generates filters derived from decorrelation for each entire positive sample, using Principle Component Analysis (PCA) method. Meanwhile, extensive pedestrian detection methods, which utilize statistic human shape to guide filters design, point out that the head-shoulder area is the most discriminative patches in typical classification stage. Inspired by above mentioned local decorrelation operation and discriminative areas that most classifiers indicate, in this paper we propose to integrate human shape priority into image patch decorrelation to generate novel filters. To be specific, we extract covariance from salient patches that contain discriminative features, instead of each entire positive sample. Furthermore, we also propose to share covariance matrix within grouping channels. Our method is efficient as it avoids extracting uninformative filters from redundant covariance of convergent patches, due to embedded prior human shape info. Experiments on INRIA and Caltech-USA public pedestrian dataset has been done to demonstrate effectiveness of our proposed methods. The result shows that our proposed method could decrease log-average miss rate with detection speed retained compared to LDCF and most non-deep methods.

Keywords: Specific patches partition, decorrelation, shared covariance, channel features, average human shape.

Received May 22, 2019; accepted May 28, 2020