Aiming at the problems of low accuracy of traditional orb-slam2 algorithm, easy loss of picture frame tracking, and the lack of dense point cloud map and octomap, the originally constructed sparse point cloud map can not be directly applied to the three-dimensional path planning of robot. Based on the traditional orb-slam2 algorithm, this paper improves the selection of key frames. Firstly, based on the traditional orb-slam2 algorithm, the comprehensive transformation factor of relative motion is added between adjacent image frames, and the inter frame feature point tracking is added to improve the accuracy of key frame selection; Then, the selected key frame is used to construct the dense point cloud map and octomap; Finally, the verification is carried out on the tum data set, and the physical test is carried out based on the real environment. The experimental results show that the improved key frame selection method can increase the positioning accuracy of orb-slam2 algorithm on the premise of ensuring the accuracy and rapidity of key frame selection, effectively alleviate the problem of easy loss of picture frame tracking, and the octree map can be directly used for robot 3D path planning.