Research Of Real-time Pedestrian Detection Technology Under Complex Background

This paper take adavantadge of phase congruency to extract pedestrian edge,combined with local energy and Log Gabor filter,compute the maxium local energy and the corresponding phase of pedestrian edge under different scale of image and filtering direction,weighted projection as the way of HOG feature generation to get the PC-HOG feature.Trained by AdaBoost algorithm to produce the strong classifier,training time decrease to 2.5 hours.Testing the classfier in the INRIA database,the detection results show that the algorithm improve pedestrian detection speed,the detection time improves 40%compares to HOG?SVM and 8%compares to HOG+AdaBoost.By down sampling the image,the detection results of PC-HOG+AdaBoost show an significant improvement in the number of pedestrian right detction and the number of false positive detection,the right detection number improves 12%and detetion time decreases 50%compares to compared to HOG+SVM,detection time can be 0.07s on a size of 320x240 image.To improve the right detection rate on small image with the usage of HOG feature,this paper use frame difference method,extracts the motion foreground binary image in video image under different threshold value,after removing the influence of background factor on pedestrian contour,trainning the HOG features with AdaBoost algorithm.The classifier can achieve a positive detection rate of more than 98%in the MIT database with a size of 64x128 image,detection time can be 4ms.When detecting and tracking pedestrian pedestrian in dynamic occasions,based on color information,back projection transformation of H color compo-nents,using Mean Shift algorithm for target tracking,which appears tarcking drift and failure.In order to improve the tracking performance,conbined with motion feature,using frame difference method and Gaussian Mixture Model to extract the motion foreground,Based on Mean Shift algorithm,merge with connected components and particle filter for target tracking,Kalman filter used to predict and correct the target loaction.Using pedestrian detector at the motion target region,detection time can be 0.3ms,pedestrian detection speed improved obviously.With the improvement of pedestrian feature under static image and pedestrian tracking algorithm in video image,this paper realize the real-time pedestrian detection under complex background.