Research On Representative Feature Subset Generation Methods For Pedestrian Detection

Pedestrian safety in Urban Traffic gets worse every year. In order to protect pedestrians and reduce the amount of traffic accidents, onboard Pedestrian Detection Systems (PDS) have become a confessed research hot spot in Intelligent Traffic System. The PDS is designed to automatically detect pedestrians in front of vehicles by the visual surveillance system and give the alarm to drivers in dangerous situations.In recent years, researchers around the world have made great progress in PDS. But most of the works focused on designing and optimalizing recognition algorithms, while research on feature extraction and selection receives less attention comparatively. Feature extraction and selection technology can be improved greatly to obtain optimal features in order to enhance detection accuracy and detection speed. Therefore, How to obtain an optimal representative feature subset is the key for the PDS improvement, and the research on representative feature subset generation is significant for theory study and valuable for application.This thesis proposed two representative feature subset generation algorithms applied in two popular detection technologies in the context of visual PDSs, which can reach a good balance on detetion speed and accuracy. The skeletons of the algorithms are shown as follows:(1) For PDS using template matching, we proposed a representative pedestrian shape feature subset generation algorithm based on an improved Local Linear Embedding. It can deal with the contradiction of pedestrian variety and the limitation of the amount of templates, and it can efficiently reach a good balance on detection accuracy and speed.(2) For PDS using classification, we proposed a representative feature subset selection algorithm based on Kernel Principle Component Analysis and Improved Genetic Search guided by a Linear SVM. It can obtain a high-discriminable and optimally representative feature subset, and it can solve successfully the difficult problem of pedestrian variety and background confusion in real time.In order to demonstrate the effectiveness of the proposed algorithms, we designed some comparative experiments. The results showed that our representative feature subsets are better than other feature sets and pedestrian detection with them can reach a well real-time and robust performance.