In large outdoor scenes, it is difficult for laser odometer to accurately register point clouds due to complex environments, resulting in a large cumulative error of motion trajectory. To solve this issue, an improved generalized iterative closest point (GICP) algorithm based on geometric and intensity constraint registration is proposed. Firstly, the average flatness and plane similarity of the point pairs to be matched are calculated by the neighborhood characteristics of the point clouds, and a geometric weight function is constructed to reduce the errors of corner points and poor corresponding point pairs during the point cloud registration process of the GICP algorithm, thereby improving the accuracy of the algorithm. Secondly, a symmetric KL (Kullback-Leibler) divergence parameter is introduced to construct intensity similarity and measure the intensity difference of point pairs to increase registration constraints. Additionally, KD-Tree is employed to accelerate the search of point cloud matching pairs to enhance the efficiency of the algorithm. The experimental results of KITTI dataset demonstrate that the average positioning error of the proposed algorithm is reduced by 58.19% and 45.64% compared with GICP and VGICP, respectively. Field experiment results show that the average positioning error of the proposed algorithm is reduced by 52.58% and 37.12% compared with GICP and VGICP, respectively, while meeting real-time requirements.