Research On Pedestrian Detection Method Based On Integrated Machine Learning

Pedestrian detection is a process that detects pedestrian from the video or image.It is an important part of computer vision. Pedestrian detection technology ha s broadapplication prospects and potential economic value in the intelligent monitoring,advanced human-computer interaction, intelligent household, etc. But due to theinfluence of various external environmental factors, such as pedestrian posture,illumination is complicated. There is still no mature method to solve this problem.Therefore, pedestrian detection applications in practice are both opportunities andchallenges.Pedestrian Detection methods can be divided into: the pedestrian detectionmethod based on the traditional approach and the pedestrian detection method basedon feature extraction and machine learning. At present, the method based on featureextraction and machine learning is more popular. The method is based on a largenumber of sample set for processing, which has certain robustness. In the view of thekey steps of feature extraction and classifier design, the present work is involved inthe following two aspects:For the image of the pedestrian detection. We proposed the new detect ionalgorithm, it combined gradient orientation histogram(HOG) feature with local binarypattern(LBP) feature and cascade Gentle AdaBoost classifier. To the problem of thelower rate of a single feature, we combined with the two features, which can describ eboth contour information and texture information. Single scale feature was proposedbefore, we proposed the multi-scale feature. At the same time, we employed integralimage method to accelerate the feature extraction. For classifier, we use the CARTtrees as weak classifier, it improved the performance of classifier. Then we used theGentle AdaBoost classifier to select some weak classifier. Finally, in order toaccelerate out of most of the background region, the cascaded Gentle AdaBoost wasobtained by training and used for testing. In addition, we illustrate two detectionmethods of scalled pyramid and a larger window. In addition, for the window fusionstrategy, we compared the integration of the mean-shift method and window-basedfusion method. Experimental results showed that our improved detection algorithmperformance by about5%more than the original detection method at FPPW=10-4.Pedestrian detection in video, a moving object detection algorithm based on combinating the interframe difference and background subtraction method wasproposed. Mixture gauss model was used for updated background. Then get thecandidate regions for mathematical morphology processing the binary image. Then,the interested region used the cascade Gentle AdaBoost classifi er for testing, in thisway, it could reduce the detection range, improved the detection speed greatly.