In this research, we have investigated the detection of drowsiness activity in dorsolateral-prefrontal cortex in three different time windows (0~3 sec, 0~4 sec and 0~5 sec) using functional near-infrared spectroscopy (fNIRS). Five drowsy subjects participated in a simulated driving task while their brain activity is monitored using fNIRS. The recorded brain activity is segmented into three windows for the acquisition of signal mean, signal slope and number of peaks as features. The data in each window is classified using linear discriminant analysis to find best window size. The results show that the best accuracy is obtained using 0~5 sec window after classification. Although the classification accuracy in 0~4 sec window is lower than in 0~5 sec window, both accuracies are suitable for brain-computer interface applications (i.e. accuracy>70%). The accuracy in 0~3 sec window is less than 70% for two subjects. For driver drowsiness detection, high accuracy with quick detection time is required, therefore we propose drowsiness detection in 0~4 sec window using fNIRS monitoring.