Abstract: Currently most of the traffic camera self-calibration algorithms are based on vanish point and geometry markers in road scene. However, for the detected multiple vanish points, there exist the unstable disadvantages and the "ill-condition" of approaches infinity. In addition, the acquired makers may not be accurate. So the practical application of traffic camera self-calibration is limited nowadays. To overcome the above problem, firstly, this paper builds a more stable self-calibration model with single vanish point based on the typical road scene;then acquires the calibration region and geometry markers dynamically, and get the optimal vanish point in diamond space;finally, use the redundant information in road scene to formulate non-linear constraints, and iterate the calibration parameters in the constraint space to get the optimal solutions. Therefore, calibration errors resulted from imprecise initial calibration conditions could be eliminated. The experiment was carried out in pan-tilt-zoom traffic camera monitoring curved road environment, where the camera angle and focal length were changed synchronously in real-time. The results show that the proposed method gets more than 95% accurate of calibration result in road scene, which is better than existing algorithms, especially it applies to the real time self-calibration of pan-tilt-zoom traffic camera.