Headlights detection is the key point for the automatic recognition of the turned-on high beams in night-driving vehicles, however, it is uneasy to identify the quite similar vehicle light targets from the video frame images. To solve this problem, an accurate algorithm for headlights region segment is proposed in this paper, which is based on the confidence intervals estimation of the vehicle width projective transformation. Firstly, the vehicles light region is extracted crudely, and then the vehicle lights come from the same vehicle are matched according to the confidence intervals of the vehicle width, finally, the high and low light region can be segmented accurately through the local projection and the precise local threshold segmentation. The experiment results show that the proposed method achieves high accuracy rate and recall rate in the headlights tests of night-driving vehicles.