A geometrical feature-based framework for pedestrian crossing recognition

Abstract

In this work, a framework for the recognition of pedestrian crossing (PC) regions based on geometrical features is proposed. A distinctive feature of a pedestrian crossing (PC) is that each end point of the horizontal strip edges at pedestrian crossings intersects with a vertical stripe width edge, which comprises two connected points (2CP). Another unique feature of PC is that the PC stripe's edges are formed in ascending parallel order. These two features are utilized to identify the PC candidate region in the PC image. Where the 2CP and parallel edge segment in sorted order is used to validate and justify the PC region. Finally, classifier support vector machine (SVM) confirms the potential PC region. Here, the features of the candidate area are extracted using the uniform rotationally invariant Local Binary Pattern (LBP). The proposed framework is tested with our own dataset, and the results reveal significant improvement over previous work.