Empirical Study of Pedestrian Detection Algorithm Based on Ensemble Learning

Abstract

Pedestrian detection is a key problem in computer vision nowadays. It has great significance of improving the quality of life in contemporary society, and it is becoming a hot research topic during recent years. As shown in previous researches, we find that a simple combination of a single feature and classifier will not perform well in pedestrian detection. Therefore, we proposed several methods based on ensemble learning to explore effective pedestrian detection ways: (1) We integrate several weak classifiers to get a strong classifier to improve the detection performance, including AdaBoost algorithm and integration of different kernel SVMs. (2) We explore the way of integrating several kinds of features to improve the detection performance. Experimental results demonstrate that different integrating methods will bring various results and they almost can improve the performance comparing to single feature or single classifier. In this paper, several kinds of effective ensemble pedestrian detection algorithms are proposed from the extensive experiments and we test our algorithms on the Bay trial and Win8 platform, finally we obtain a promising result.