Pedestrian Crossing Detection and Recognition Based on Two Connected Point and Uniform Local Binary Pattern

Abstract

A geometrical feature base pedestrian crossing (PC) region detection and recognition framework is proposed in this work. A unique feature is that each end point of the horizontal strip edges of pedestrian crossings is intersected with a vertical stripe width edge. That makes up two connected point (2CP). Another feature is that the edges of the PC stripe are formed in ascending parallel order. By utilizing these two features the PC candidate region is detected from the PC image. Where, the PC region is validated and justified by using 2CP and arranging the horizontal parallel edge segment in sorted order respectively. Finally, the potential PC region is confirmed by a classifier i.e., support vector machine (SVM). Here, the features of the candidate area are extracted using a rotationally invariant uniform Local Binary Pattern (LBP). The proposed method is tested with our own dataset and results reveal significant improvement with respect to the existing works.