On-Road Vehicle Detection Based On Gabor Filters

Safety Driving Assist is an important study content in the field of the IntelligentTransportation Systems, it mostly solves the problems of the traffic safety, and italso can release the traffic jam and environmental pollution which is persecutedtransport domain.The Vehicle Detection is one most important research in ITS, it is the most keystep in Safety Driving Assist,Self-motion Navigation,Traffic Monitoring,Scheduling Monitoring and Accident Detection. Generally speaking, On-roadvehicle detection consists of two main steps: Hypothesis Generation andHypothesis Verification. During the Hypothesis Generation step, the location of oneor more vehicles in an image is hypothesized. In the Hypothesis Verification step,the true existence of vehicles at the hypothesized locations is tested.Accordingly in my papers, the algorithm of the vehicle detection includepreprocessing in the gray image, location the Area Of Interest (AOI) whichpotential contains the vehicles, feature extraction using the Gabor filters,classification by using performed Support vector machine to verification whether ovehicles exit. It can be described as following: Hypothesis Generation, establish the Area Of Interest. Using the Gabor filters to extract the vehicle features as the input of theSVM. Verification whether vehicles exit by using performed SVM.In this paper, according to the features of the On-road vehicles in the grayimage, we bring forward the method to location the AOI as following: In gray image, we use the gray equalize algorithm to add the dynamicrange of the pixel gray scale, so it can enhancement the whole contrast of the image,force the image to have relatively contrast and more details sharply. We use the one-dimensional maximum entropy threshold algorithm tosegment the image. For removing some useless interference, we introduce theregion growing and threshold area algorithm, after these we binarization thesegmented image.After step two, we use the features of the vehicle in the gray image (suchas the under-vehicle shadow, the symmetry of the vehicle) to establish the AOI, thedetail method will be explained in the text.In the Hypothesis Verification, we separate it into feature extraction andclassification identify. Because of the superiority of the Gabor filters in the featureextraction, especially succeed apply in the other domain (such as Face Recognition,Handwritten Chinese Character Recognition, Speech Recognition), we extract thefeatures of the vehicle by using the Gabor filters.The feature extraction algorithm requires both strong robustness and strongregimentation representation capability, so in this paper we present the featureweighting, that is, the extracted features are weighted according to theirneighboring features degree of dispersion. To reinforcement the effect of thefeatures whose degree of dispersion is relatively small and weaken the effect of thefeatures whose degree of dispersion is relatively biggish. Moreover, we make usefully of the information of the location and statistic in the sample images, so thatwe can improve the robustness and accuracy efficiency.The vehicle verification is really a problem of the binary pattern recognitionessentially, it is that the classification of the vehicle and background. In view of theSupport Vector Machine (SVM) is more usefully in the pattern recognition recently,we adopt the SVM to classify the vehicle and background.Although the SVM is no sensitivity with different methods, the training result isstill different when the kernel function and parameters. So we select the radialgauss function as the kernel, select the Sequential Minimal Optimization(SMO)algorithm bring forward by Platt, it can be come true easy.The process of the recognition embody as following:We collect sample images (include rear vehicle views and background) indifferent time, we scale the images to 64*64 when we extract the features use theGabor filters. And the features vectors are the input of the SVM.Select avail kernel function (radial gauss function in our paper) and thetraining algorithm (SMO), training the SVM with the input features.We use the algorithm that has been introduced in the front to verify theAOI. And then extract the features by Gabor filters that has been expatiated in thethird chapter of the paper. Finally we verify the object whether is vehicle by usingthe training SVM.The software application system is developed using Visual C++. The functionof the application is proved to be reliable. The experimental results prove thevalidity of the algorithms mentioned in this paper.