Research On Fast Pedestrian Detection Algorithm In Complex Background

Pedestrian detection is used widely in computer vision system, such asintelligent video-surveillance, automatic drive, multimedia retrieval and so on. Inaddition, pedestrian detection is also the basis of many subsequent processingsuch as scene analysis, motion analysis, behavior understanding and so on.Because pedestrian is nonrigid, the pedestrian’s posture, appearance, perspectiveand circumstances vary greatly, a lot of existing methods and models are eithertoo simple to achieve detection precision or too complicated to meet real-timerequirement.According to the issue of pedestrian detection in complex background, thispaper presents an Improved Center-Symmetric CENTRIST (ICS_CENTRIST)feature from the view of the pedestrian edge information, which is based on theCensus Transform Histogram (CENTRIST) statistical feature descriptorcombining with Center-Symmetric properties of Center-Symmetric Local BinaryPattern (CS_LBP) feature. ICS_CENTRIST feature describes pedestrian’s edgecontour information by histogram of blocks and encodes edge image with only32dimensions, which has characters of simple calculation and powerful descriptionability.ICS_CENTRIST feature is based on traditional uniform block generatingmethod, which cannot use any prior knowledge and includes a lot of redundantinformation in extracted feature vectors. This paper uses stable block selectingmethod based on Partial Least Squares (PLS). First one major projection vectoris selected for every block. The projected value is used as the feature value of theblock, and then PLS model is built for all sample block feature values. A scorewill be calculated out for each block according to the model parameters, whichrepresents predictive and separating power of this block in the PLS model. A setnumber of blocks with high scores will be selected as the area to extract featurein detecting window.Three cascaded classifier are used for pedestrian detection. In the first stage,the linear SVM based on auxiliary integral image is used for excluding mostnon-pedestrian area quickly. During the second and third stages, the first12and 21blocks with most strong distinguishing ability chosen by PLS method areaccepted respectively, ICS_CENTRIST features are extracted and thenHistogram Intersection Kernel SVM (HIK-SVM) is used for accurate detecting.Experimental results show that the algorithm presented in this paper can getbetter detection results in complex background, and the detection speed isaverage50ms for447×358images, which is improved by50%and90%compared with the CENTRIST fast detecting method and HOG algorithmrespectively and can meet the real-time requirement.