Research Of Pedestiran Detection Method Based On Multi-features

Usually, visual video surveillance system includes a huge amount of cameras, and eachcamera monitors a relatively fixed area. Typically, after a certain time, people monitordifferent views by switching the perspectives of different locations or record regionalmonitoring information as a means of investigation after certain incidents. The combinationof different technologies such as pedestrian detection, behavior analysis and videosurveillance, can provide a solution for the real-time intelligent safety monitoring system.The research of pedestrian detection meets the urgent needs of real-time alarm in theintelligent monitoring area. Pedestrian detection and subsequent abnormal behaviorrecognition are very important parts of the real-time monitoring field.Pedestrian detection is currently an important part of the computer vision and patternrecognition fields. A complete pedestrian detection system consists of two parts: the firstpart is extracting the interested area, usually by moving object detection; the second part isscanning a single image to find the pedestrians in the interested area.Moving object detection is separating the monitored foreground object from background.It is an important processing step in the visual human-computer interaction，image analysisand intelligent monitoring fields. Through moving object detection, we can get the movementinformation and extract the moving pedestrian or object, so that to simplify the follow-uptracking and analysis. At common, the object we focused in is always in the object detectionarea. At present, there are several methods for this problem including Gaussian mixturebackground model, the difference method and the optical flow method.In this paper, we describe several commonly-used moving object detection method suchas Gaussian mixture background model, difference method and optical flow method. Thispaper focuses on the Gaussian mixture background model. It can efficiently deal withmultimodal distributions caused by shadows, swaying trees and other knotty problems of thereal world. However, the method suffers from foreground objects bending into thebackground too fast. In addition, it can’t deal with the problem of slow-moving objects. Inthis paper, an improved method is proposed to deal with the above problems. Theexperiment result shows the method works better than the typical Gaussian mixture model.The main idea of current pedestrian detection strategy is to extract pedestrian features,then train classifier using training samples, finally use the classifier to detect pedestrians in aspecific region. Meanwhile, some researchers use the method of calculating the similaritydistance between pedestrians and non-pedestrians. The simple features extracted forpedestrian detection are aspect ratio, the object’s relative movement speed and so on. Some complex features can be haar-like feature, hog, sift, shapelet and so on. Haar-like features isused for face detection at first, and achieved good performance, and then it is introduced intopedestrian detection. The main idea of hog feature is the appearance and shape of an imagecan be fully described by gradient or edge of the density distribution. Sift is an algorithm toextract the local features, the method can find extreme points in scale space and extractionlocation, scale, rotation invariant. Choosing efficient features for pedestrian detection is ahard problem.In this paper, we adopt the combination of histogram of oriented gradients (hog) andGabor as the pedestrian detection feature description, and use the linear support vectormachine to accomplish the problem of pedestrian detection. The combination of gradientorientation histogram and Gabor features can describe the features of pedestrian better. Thetrained pedestrian classifier model has better classification results with higher accuracy.