Aiming at the problem that many feature extraction algorithms for visual odometry have large deviation of motion information due to uneven corner distribution, a corner extraction method based on Gaussian Pyramid is proposed. Firstly, the algorithm adopted the Gaussian Pyramid algorithm to compress the image scale, compressed the texture in the area of rich texture and aggregated texture in the area of sparse texture. Then, the small-scale top diagram was obtained and the Shi-Tomasi algorithm was used to extract the corners of the top graph for rough feature location. The information was mapped back to the original graph with rich details for accurate corner positioning, and the feature information of the picture was obtained. Finally, the Pyramid LK optical flow method was used to track the corner points, and the motion information is obtained according to the geometric constraint model of camera polar line. In this paper, the KITTI data set was used to compare with the original Shi-Tomasi, Harris, Fast Algorithms. The results show that the algorithm can effectively improve the uniformity of corner distribution and improve the accuracy of motion information recovery of visual odometry.