Abstract: Driving under stress status has a great impact on drive safety and even causes traffic accidents in severe cases. It is important to develop appropriate approaches for early detecting the driving stress and then improving the awareness and performance of drivers. Due to its strong objectivity, robustness and real-time property, the physiological signal analysis has been validated as an effective approach for reflecting physiological conditions of human operators. In order to accurately detect the stress status of drivers, multi-modal features are extracted from time, spectral, wavelet, and nonlinear domains based on physiological signals, i.e., electrocardiogram, galvanic skin response and respiration of drivers. And then, a hybrid algorithm based on the multi-filters (MF) and tabu search (TS) is proposed in this work to select effective features, since there are many redundant features and invalid features in the feature set. Firstly, in order to effectively reduce the feature dimension, the original features are sorted and filtered according to the comprehensive score calculated by multiple filtering algorithms. Three different filter methods including Chi test, Wilcoxon test and manual information are employed to generate a new feature subset, which had smaller feature dimensional. And then, the tabu search algorithm is adopted to further select the optimal feature combination from the optimal feature subset. Finally, three different drive stress levels are classified by means of support vector machine (SVM). Some widely-used feature selection methods, including genetic algorithm (GA) and random tabu search algorithm, are also implemented for comparisons. The proposed hybrid algorithm not only eliminates redundant information in high dimensional eigenvectors effectively, but also improves the classification accuracy.