A heuristic SVM based pedestrian detection approach employing shape and texture descriptors

Abstract

Pedestrian detection is a vital issue in various computer vision applications such as smart security system, driverless car, smart traffic management system and so forth. However, the issue of low detection accuracy and high computational complexity still makes a prompt topic of research. In the current scenario, Histogram of Oriented Gradients (HOG) with linear Support Vector Machine (SVM) is considered to be the most discriminative detector and has been adopted in various advance systems. In this paper, a novel method for pedestrian detection is proposed with the objective of improving the detection accuracy, precision and other metrics values. The proposed approach combines Histogram of Significant Gradients (HSG) and Non Redundant Uniform Local Binary Pattern (NRULBP) to generate a competent descriptor to be used in our detection model. The proposed approach is used in conjunction with various classifiers and the linear SVM classifier is found to provide better metric values over others. Different datasets like INRIA, TUD-brussels-motion pairs and ETH are utilized for performing experiments and to obtain detection results. Experimental results show that the proposed descriptor outperforms HSG by 2.59%, 8.97%, 8.5% and NRULBP by 3.19%, 39.55%, 19.66% in terms of detection accuracy, precision and F1 score respectively.