Pedestrian Detection Algorithm Using Multiple Features And Improved Houghforest

Pedestrian detection is a branch of object detection in the domain of patternrecognition, which is aimed at locating and classifying the pedestrian fromsequential video frame and single frame. Meanwhile, it is one of the mostactive and challenging tasks in the field of computer vision. Pedestriandetection has been attaching the attention of researchers day by day justbecause of its broad application prospect in the realm of Driver AssistanceSystems, Advanced human-computer interaction and Smart Monitoring and soon. However, due to the characteristics of pedestrians themselves andsurroundings, such as variable appearance and the wide range of poses,clothing, differences between individuals, lens Angle and illumination and soon, there exist fixed difficulties to accurate detection. Recently the mainstream research strategy of pedestrian detection algorithm is to describepedestrians by exacting their characteristics, and then train the collectedfeatures. Exactly the fittest features by computer learning algorithm, we canform classifier then detecting humans in images.This paper is aimed at pedestrian detection problem in the context of nature,setting up a new pedestrian detection system. The main innovations are asfollows:（1）In the matter of feature extraction: Starting from macroscopichuman eye vision, this paper presents a descriptor which synthesizes themultiple-features of pedestrian local gradient information, texture informationand significant color information. The test has proved that the feature set has afavorable descriptive power toward pedestrians.（2）In the matter of classifier:Owing to descriptor has chosen multi-feature power, this paper choosesHough forest algorithm which can melt into more identifying information toset up classifier. Hough forest can learn efficiently and choose the bestgradient, texture, color feature of characteristic set, then combine featureextraction and classifier training, presents a new pedestrian detection algorithm.（3）We found that the vote pattern of Hough forest has existedlimitations. Consequently this paper has improved that, presents a regionalweighted voting template which can improve detection precision.This article implements pedestrian training and detection base on thesoftware of VS2005and OpenCV. Experimental results show that when errordetection rate is10-4FPPW, the detection rate is90.12%, ROC curve issuperior to HOG+SVM arithmetic and original Hough forest arithmetic.