The Research Of Pedestrian Detection Based On Multiple Features And Adaboost

Recent years, Pedestrian Detection has become one of the most popular research focus in thefiled of computer vision. as an important direction of target detection, pedestrian detectiontechnology have a wide range of applications in intelligent surveillance, intelligent transportationand driver assistance. As the main object of the video scene, the behavior analysis andunderstanding of pedestrian is obviously an important task of computer vision.At the same time,pedestrian detection as the important prerequisite and basis for the analysis of pedestrian behavior,so accomplish the pedestrians detection accurately and efficiently gets more and more attention.However, due to the complexity of the background interference, illumination changes and occlusionwhen practical application, although pedestrian detection technology has been widely studied, butstill now does not have a generic and robust pedestrian detection algorithm could meet the accuracyand speed requirements of practical applications.Detection methods of statistical learning is mainstream in recent years, pedestrian detectionmethod based on the sample from the library contains a large number of samples in the positive andnegative pedestrians, extract pedestrian feature, learning to generate statistical models ofpedestrians, translate pedestrian detection problem into a machine learning classification problem.In2005,Dalal proposed a very good feature which we called gradient direction histogram (HOG)feature has a perfect description of the human body, using support vector machine learningclassifier generation, achieve a very good detection results, the algorithm has also become one ofthe classic algorithm in pedestrian detection field. However, the dimension of HOG feature is toohigh, and the use of support vector machines for training and testing the feature is a verytime-consuming process,so it is difficult to meet the needs of real-time monitoring. The Adaboostcascade object detection method proposed by Viola et al has a simple frame structure, with a highdetection speed.Therefore, this paper combine HOG features and Adaboost cascade framework to detect human,and make improvements in the following areas:a)for a better use of the multi-feature learning abilityof Adaboost algorithm,this paper combine HOG features and Haar-Like feature to build hybridfeature library, take advantage of HOG ability to describe the characteristics of the body contourand Haar-Like feature to describe the local details, achieve a fuller expression of the human bodymodel; b) To further enhance the diversity of HOG features, using variable block size HOG feature extraction method，for the problem of high dimensionality of HOG features, this paper use fisherlinear discriminant criterion weighted HOG features for dimension reduction, this help HOG featurea better integration to Adaboost algorithm and improve the detection rate;c)this paper use RealAdaboost algorithm instead of using Discrete Adaboost algorithm, the traditional thresholdclassifiers will be replaced by higher classification accuracy based on a lookup table type.By the improvement in above aspects, the validation results in INRIAPerson sample libraryshowed that: the proposed multi-feature-based Adaboost cascade pedestrian detection algorithmeffectively improve the accuracy of pedestrian detection and reduces the training and testingtime.In real scene detection rate can reach90%, while in320*240resolution on the relative speedof the original image detection algorithm has been greatly improved and could meet the basic needsof real-time detection.