Research On Pedestrian Detection Based On Multiple Features And Cascade Detection Method

Pedestrian detection technology has always been a central issue in the field of computer vision. With the development of computer science technology, pedestrian detection technology has been gradually applied to reality, such as Intelligent monitoring system, Automotive auxiliary system, Intelligent Robot and so on. The demand of social development attracts more and more researchers who are dedicated to the study of pedestrian detection technology. HOG feature which was proposed for pedestrian detection in 2005 is the breakthrough of the field of pedestrian detection.Numerous methods are used to pedestrian detection in the last decade and they achieve good performance. There are many difficulties on pedestrian detection.(1)As pedestrians are non-grid targets, their appearance, postures, etc. will change at any time and influence accuracy.(2)How to get rid of the complicated background interference becomes the difficulty of study. Complicated background could lead to miss detection and false detection.(3)The phenomenon of pedestrian shelter will appear in the crowded scene while many people exist in the crowded scene. The reason why we try to eliminate the effect of pedestrian shelter is pedestrian shelter itself influence the complexity of pedestrian detection.This paper adopts cascade detection method, main contents are as follows:(1)Proposed improved HOG feature for pedestrian detectionTraditional HOG feature method for detection just extracts single feature while single feature can’t describe more pedestrian target feature information. In consideration of numerous features which could describe target, this paper proposes combination of HOG feature and Entropy. The definition of Entropy originates from thermodynamics.However, when Entropy is applied to image process, Entropy could describe image grain feature. Experiment takes advantage of improved HOG feature and slide window detection method. As a result, we can get a series of preliminary rectangle of interests.Due to the limited detection ability, result will contain false rectangle of interests. While it doesn’t matter, the result we get will be precisely distinguished.(2)Proposed collection channel feature for target detectionCollection channel feature consists of color feature, gradient magnitude feature and HOG. Multi-feature integration would improve the accuracy of the pedestrian detection to a certain extent. This paper proposes fast feature extraction method which can quickly compute variable scale image feature. Fast feature extraction mainly extract image feature in the current scale and then get approximate image feature in adjacent scale. In contrast with traditional feature extraction method, fast feature extraction save a lot of computation time. The feature will be trained by Adaboost algorithm. After the rough detection step, precise detection will get good performance in this step.In this paper, the databases used in the experiments are open source public pedestrian databases. Experimental platform is Intel Core Duo processor, 8G memory,Matlab 2014 a. The method of combination multi-feature and cascade detection will improve the detection results.