Abstract: The real-time performance of vehicle detection is very important in an intelligent transportation system. Conventional Histogram of Oriented Gradients (HOG) method has problems of lots of dimensions of feature vector and huge calculation. Therefore, this paper studies the characteristics of gradient distribution of vehicles and the influence of feature’s dimension to Support Vector Machine (SVM)’s time performance. Therefore， the paper proposes a vehicle detection method, which combines local gradient vector’s mean and scatter matrix with SVM. First of all， the sampled image is divided into a number of blocks uniformly. Then the gradient vector’s mean and scatter matrix are calculated as feature vectors in each block. At last, the classification and identification are performed by SVM， which further reduces the calculation by variable step size. The experimental results show that the method’s accuracy is equal to conventional method, but the average recognition time is reduced to 51% of conventional method.

Key words: gradient vector’s mean, scatter matrix, Histogram of Oriented Gradients (HOG), vehicle detection