A calibration method based on the principle of a non parallel ranging model was presented to obtain the parameters of the binocular stereo vision device. The ranging error caused by the non parallel optical axis and the measurement error of optical center distance were decreased. In order to slove the problem that there were lots of error matching pixels and target recognition rate was low under low illumination in the binocular stereo vision, the threshold selection method which combined the maximum entropy and change of light intensity was proposed to extract significant object, the recognition rate was improved. Finally, the significant target recognition system was build. The results show that the binocular stereo vision device, utilizing the method of camera calibration and threshold selection proposed in this paper, has a higher target recognition rate and ranging precision under low illumination.