The Research On The Algorithm Of Pedestrian Detection Based On Infrared Image

The technologies of pedestrian detection have become one of the most activeresearch topics in pattern recognition and computer vision. The images collected byvisible light cameras are always blurry when the light condition is poor at night,which can influence the effect of pedestrian detection. While the infrared camerashave good night vision and a strong ability to different light environment. So theapproach to detect pedestrian based on images taken with infrared cameras hasbecome a kind of effective method to solve this problem. However, people not onlyhave rigid object features but also have flexible object features. And people’sappearance is vulnerable to their posture, the influence of perspective and occlusion.Thus, the technologies of pedestrian detection based on the infrared images under theenvironment of night are still a challenging research topic.Aiming at the vehicle scene at night, this paper mainly studies the pedestriandetection problems related to feature extraction methods combined machine learningmodels. The main research contents are as follows: feature extraction, the pedestriandetection algorithm, real-time detection methods and so on. The main works are asfollows:Firstly, for the problem of feature extraction in the field of pedestrian detectionwith the infrared images, this paper put forward a kind of descriptor DBHOI(Different Bins Histogram of Intensity) which is more capable of descriptive ability.By studying HOI (Histogram of Intensity) descriptor which characterizes the featureof luminance information in the infrared images and its construction process, wediscover that HOI does not make full use of the characteristics of human beings andthe distribution of luminance information. Taking above disadvantage intoconsideration, this paper come up with a new descriptor, DBHOI feature descriptor,which can code the pedestrian’s brightness information distribution. Finally, adetailed optimization process involved in the construction process of DBHOI featuredescriptor is given.Secondly, this paper presents a novel pedestrian detection training framework forinfrared images and uses an approximate model to accelerate the speed of detection.First of all, a novel integral channel feature is constituted by combining DBHOIdescriptor and HOG (Histogram of Gradient) descriptor. The constituted integral channel feature can be used to describe the pedestrian’s luminance information andmarginal information in the infrared images. Then, based on the Adaboost trainingframework, the two layer decision tree model is used as weak classifier, and a cascadeclassifier with a strong classifying ability can be achieved by using a large amount ofdata. Finally, in order to meet the requirement of real-time detection, this paper alsoapplies a new method which combines the image pyramid model and the classifierpyramid model to calculate the characteristics between adjacent layers approximately,thus reducing the overall detecting time.In summary, on the basis of analysis of existing researches, this paper studiesfeature extraction, the pedestrian detection algorithm, real-time detection methods.The experiment results have shown the presented DBHOI descriptor and Adaboosttraining framework perform good and robust. This study contributes to thedevelopment of intelligent vehicle auxiliary driving technology.