Research On Pedestrian Detection Based On Feature Fusion And Ensemble SVM

As one of important branches in the area of computer vision,Pedestrian detection has been widely applied to some areas such as controlling system of intelligent transportation;auxiliary system of superior driving and the gesture evaluation.At the sometime faced with a series of problems such as changing appearance and complex background environment,pedestrian detection has aroused general interest with the highly-research value and challenging.Based on the method of statistic learning,this paper lays emphasis on the method in view of HOG feature and SVM classifier,and proposes a pedestrian detection method based on feature fusion and Ensemble SVM classifier to solve the problem of poor accuracy caused by single feature and classifier overfitting.The main work as follows:1.Since the accuracy is low due to the imbalance of positive and negative samples in SVM pedestrian detection method which adopt under-sampling method.For solve this problem,this paper proposes purpose a pedestrian detection method of Ensemble SVM classifier which combining the two training mechanism of the under-sampling SVM classifier and the divide and conquer strategy of EasyEnsemble: Firstly,selecting negative sample as initial training sample randomly that is divided into multiple sub-negative sample sets equalizing the positive sample set,building balanced sub-training sets and linear assembling EasyEnsemble SVM.Then negative sample is classified and judged by using EasyEnsemble SVM,making misjudged sample as hard sample,building balanced sub-training sets again and training sub-classifier,combining which with EasyEnsemble SVM to get the Ensemble SVM classifier.The method was performed on the INRIA data set to verify the performance of the pedestrian detection algorithm of the Ensemble SVM.2.As for the problem that the single HOG feature cannot represent the pedestrian's information fully and handle the pedestrian occlusion accurately,and a pedestrian detection method based on feature fusion is proposed,which is fusing the feature of HOG and Gabor to train the Ensemble SVM classifier.Experiments were performed on the INRIA data set which adding the casual noise block and actual occlusionseparately,and the results show that the fusion of HOG features and Gabor features was helpful to the detection accuracy.As the experiments show,the use of Ensemble SVM classifier can well solve the problem of overfitting caused by the imbalance of positive and negative samples data.And it also solves the problem of occlusion in the pedestrian images by the means of feature fusion.