An SVM parameter learning algorithm scalable on large data size for driver fatigue detection

Abstract

Support vector machine (SVM) is a classification model that separates instances by maximizing their distance to a classifying hyper-plane. SVM has been applied successfully to solve a wide range of application problems. However the effectiveness of SVM largely depends on the parameters used by its kernel functions. For a SVM with the radial basis function (RBF) being the kernel function, two parameters control the SVM training, c and γ. Traditionally, grid-search technique is applied to selecting the proper values of the two parameters. The grid-search method is computationally expensive when the size of training samples is large. In this paper, we present a parameter learning algorithm, Distributed Learning and Searching (DL&S). It is composed of two stages: distributed searching for significant parameters and finding optimal parameters fit for all training data. We applied the DL&S algorithm to solve an important automotive safety problem, driver fatigue detection. We present a driver fatigue detection system using a SVM trained on driver performance data, lane position, lane heading, and lateral distance. We apply the DL&S algorithm to select optimal parameters and use them to train a SVM for driver fatigue detection. Our experimental results show that the SVM generated by the optimal parameters selected by the DL&S algorithm can perform nearly as well as the SVM generated by the parameter pair found by the grid-search, and, more importantly, the DL&S algorithm consumed only 7.5% of the computational cost needed by grid-search.