Research On Pedestrian And Vehicle Detection Algorithm In Intelligent Video Surveillance

With the rapid development of computer vision and image processing technology in recentyears, video surveillance industry rises quickly and how to improve the detection speed andaccuracy of pedestrian and vehicle in video has become a serious problem to be solved. Thetraditional video surveillance system highly depends on manual interpretation of information,which greatly limits the performance of initiative and real-time systems. At present, there are alot of researches in pedestrian and vehicle detection at home and abroad in intelligent videosurveillance. The surveillance system based on image processing and pattern recognitiontechnology could quickly evaluate the scene and reduce the false positive rate and false negativerate caused by visual fatigue.A set of pedestrian and vehicle detection system in intelligent video surveillance wasdesigned by accelerating the background update speed and improving the detection accuracy inthis paper. The main study of this paper included these aspects:(1) Background modeling and update by Gaussian mixture modeling. The image sequencewas divided into blocks and marked the regions of interest after preprocessing. The backgroundwas updated according to gray value changes of part pixel of blocks in region of interest. It couldquickly and efficiently update the background by the amount reduction of data processing,accelerating the speed of background modeling.(2) Pedestrian feature direction based on histogram operator of oriented gradients. It usesgradient direction histogram feature of local area of images to describe pedestrian contour andmotion information in this module, achieving a rich feature set. The normalization of imageGamma and color space could effectively reduce local shadows of images and changes of light,while gradient calculation could capture texture information and weaken the influence of light,and projection in gradient direction could preserve the weak sensitivity of posture andappearance of the pedestrian.(3) Vehicle feature extraction based on wavelet analysis. The sampled signal wasdecomposed by wavelet transform to obtain the high and low frequency components. Because the wavelet transform has the characteristic of multi-resolution analysis, it adjusted the resolutionto achieve optimal feature extraction according to the size of texture feature and contrast ratio. Inaddition, it could effectively extract and analyze the local signal by frequency spectrum,amplitude and continuity of the signal.(4) The feature set was trained and classified by combination of SVM and AdaBoostclassification algorithm. It could efficiently control the accuracy of feature classification throughthe control of accuracy rate of iteration and amount of weak classifiers step by step.