Traditional methods of license character extraction cannot meet the requirements of recog-nition accuracy and speed rendered by the video vehicular detection system. Therefore, a license plate localization method based on multi-scale edge detection and a character segmentation algo-rithm based on Markov random field model is presented. Results of experiments demonstrate that the method yields more accurate license character extraction in contrast to traditional localization method based on edge detection by difference operator and character segmentation based on threshold. The accuracy increases from 90% to 94% under preferable illumination, while under poor condition, it increases more than 5%. When the two improved algorithms are used, the accuracy and speed of automatic license recognition meet the system's requirement even under the noisy cir-cumstance or uneven illumination.