Pedestrian Detection & Identification Based On Video And Image

Pedestrian detection and identification in videos and images are always two important tasks of object detection, which own broad applications such as transportation, commerce and security. There are several uncontrollable factors that affect previous works to get comfortable detection results in images, such as clothing, appearance and posture of pedestrians,also including background and illumination in the scene. Pedestrians, as detecting object, widely present in various types of surveillance scene. Based on comfortable detecting results of pedestrian detection, one can process tasks just like tracking, appearance and behavior recognition. A wide range of commercial value and the difficult complexity are the two factors lead scientists to have depth study on pedestrian detection, including seeking a good description of image features and searching a novel training methods.Pedestrian detection technology based on machine learning methods is facing several major problems to achieve real‐time and robust pedestrian detection in some scenes. Firstly, it takes a lot of manual costs for cutting out pedestrian samples from specific monitoring. And one can’t obtain a desired effect in peculiar scene if one only use other pedestrian samples datasets without making samples from new scenes. Secondly, it is a difficult problem to obtain a robust and efficient pedestrian detector and automatically adapt all types of scenarios.This thesis proposes generic machine learning methods in the field of computer vision to improve pedestrian detection. The main works are as follows.1) Pedestrian detection is separated into two modules: the feature extraction and training.2) About the feature extraction, according to researching and learning method, classification is based on learning with feature extraction and extraction. On the other hand, two types of mainstream statistical training methods are introduced in the feature training. Particularly HOG feature and CENTRIST feature are compared based on SVM to achieve real‐time pedestrian detecting.3) A method of automatically adaptive scene pedestrian detection for generating pedestrian detector is proposed with adaptive scenes. Firstly, CENTRIST feature is used to achieve real‐time detection. An novel method of automatically adaptive scenarios is proposed to detect pedestrian.