Optimal Feature Selection for Pedestrian Detection Based on Logistic Regression Analysis

Abstract

This paper describes a pedestrian detection method using feature selection based on logistic regression analysis. As the parent features, Haar-like and Histograms of Oriented Gradients (HOG) features are used manually. For the statistical analysis, stepwise forward selection, backward elimination, and Least Absolute Shrinkage and Selection Operator (LASSO) methods are applied to our Logistic Regression Model for Pedestrian Detection (LRMPD). The experimental results shows that the average of 48.5% of a full model are selected for LRMPD and this classifier shows performance of up to 95% for detection rate with an approximately 10% false positive rate. Processing time for one test image is about 1.22ms.