Pedestrian detection based on bidirectional local template patterns

In this paper, a novel feature named bidirectional local template patterns (B-LTP) is proposed to achieve pedestrian detection in still image. This feature is a combined and modified version of histogram of template (HOT) [1] and center-symmetric local binary patterns (CS-LBP) [2]. For each pixel, four templates are defined, each of which contains the pixel itself and two of its neighboring center-symmetric pixels. For each template, not only the relationships between three pixels according to the template, but also information of two directions are calculated in our feature, which makes it more discriminative. Moreover, the feature length of B-LTP is very short, which costs less computational workload and memory consumption. Experimental results on INRIA dataset show that both the speed and detection rate of our proposed B-LTP feature outperform other features such as histogram of orientated gradient (HOG) [3], HOT and Covariance Matrix (COV) [4].