Abstract: When the average gap between fronts of vehicles is relatively narrow during a heavy trafic, the traditional histogram matching method based on directional gradient histogram (HOG), which is based on intercepting the whole vehicle, is difficult to obtain the accurate detection area and this method has higher omission rate and false detection rate. For improving the accuracy, a new method is proposed based on the characteristic value extracting from the partial HOG. The principles of this method are as follows: at first, mid-value filtering is applied for preprocessing graph sample;then a virtual detection line is drawn in the selected region of the sample graph and acts as the sources of the region of interest(ROI) that is used for extracting the partial HOG feature vectors of the gray level graphs; at last, support vector machines (SVM) are used to train the local HOG feature vectors, and the two states of vehicle within detection line and out of detection line are classified and counted.To overcome the noise problem existed in the output of SVM,the proposed method employs a strategy combining generating a detection alignments and dual confirmation module. To improve the delicacy, the variable that describes the shadow of the vehicle is added into our SVM model. Compared to the traditional method adopting HOG, based on the whole outline of the vehicles, the method proposed in this article has the advantages of high detection rate, good real-time performance and high sensitivity. Especially when the average gap between fronts of vehicles is relatively narrow during a heavy trafic our method is more efficient.

Key words: virtual detection line, local orientation gradient histogram feature, support vector machine, vehicle detection