This paper presents a method for automatic sleep arousal detection, which is comprised of three parts: signal preprocessing, features extraction and model training. The method selects 10 channels of arousal associated physiological signals (or data) for research, including 6 channels of electroencephalography (EEG) signals, 3 channels of electromyography (EMG) signals and one channel of oxygen saturation (SaO2) data, with the fixed sample rate of 200 Hz for each channel. For the preprocessing stage, EEG signals are eliminated electrocardiography (ECG) induced artifacts first. EMG signals are then passed through a high-pass filter to reduce the low frequency noise. For the features extraction stage, total 112 features for every 3-second window are selected and prepared for the training, mainly containing EEG frequency domain features of delta, theta, alpha, sigma and >16Hz band, and also EMG time domain features. Finally for the model training stage, we choose support vector machines (SVM) with a 3-second signal window slide without overlap, and use grid search to get the best training parameters. For SVM we choose a linear kernel because of its better performance than Radial Basis Function (RBF). As a result, we get the AUC of 0.8243 by 5-fold cross validation of total training data.