Precise localization is an essential issue for autonomous driving systems. 2D LiDAR, as a high-precision sensor, is widely used in various indoor localization systems. However, in the outdoor environment, the existence of a large number of dynamic targets makes the matching of adjacent point clouds particularly difficult. Moreover, the point cloud captured by 2D LiDAR is sparse, leading to the localization accuracy of 2D LiDAR in the outdoor environment being very low or even unable to achieve localization. Therefore, a localization system fusing stereo vision and 2D LiDAR is proposed in this paper. Stereo vision is used to calculate the relative pose, so as to fuse the 2D LiDAR data in a local time window into a local submap. Dempster-Shafer evidence theory is used to fuse temporal information in local submap to eliminate noise caused by dynamic targets. Finally, the ICA-based image matching method is used to match the local submap with a pre-constructed global prior map to eliminate the cumulative error of the stereo odometry. The experimental results on KITTI dataset show that precise localization can be achieved only by using a low-cost stereo camera and 2D LiDAR. Compared with the odometry of ORB-SLAM2, the proposed localization system improves the localization performance by 37.9%. Its performance is also comparable to that of the localization system based on 3D LiDAR, and even superior to the latter on some sequences.