Pyramid Center-symmetric Local 1 Binary / Trinary Patterns for Pedestrian 2 Detection

Yongbin Zheng,Chunhua Shen,Richard Hartley,Xinsheng Huang

2010-01-01

Abstract:Detecting pedestrians in images plays a very important role 6 in many computer vision applications such as video surveillance, smart 7 cars and robotics. Feature extraction is the key for this task. Promis8 ing features should be discriminative, robust and easy to compute. This 9 paper presents a novel and efficient feature, termed pyramid center10 symmetric local binary\ternary patterns (pyramid CS-LBP\LTP), for 11 pedestrian detection. The CS-LBP feature combines the desirable prop12 erties of the standard LBP, which can be viewed as texture-based fea13 tures, and the gradient based feature. Moreover, the pyramid CS-LBP\LTP 14 is easy-to-implement and computationally efficient. Experiments on the 15 INRIA pedestrian dataset show that the proposed feature outperforms 16 Histograms of Oriented Gradients (HOG) feature and comparable with 17 the start-of-the-art pyramid HOG(PHOG) features, when using the In18 tersection Kernel SVM classifier. Our experiments also show that the 19 combination of our pyramid LBP feature and the PHOG feature could 20 improve the detection performance significantly. 21