Point cloud registration is a key step in 3D data processing. Aiming at the problem of low registration efficiency due to the weak representativeness and descriptiveness of feature points in the registration process, a point cloud registration method based on the improved 3D scale-invariant features (3DSIFT) algorithm is put forward in this paper. Firstly, the feature points extracted by the 3DSIFT algorithm are streamlined by combining the information entropy theory, and the representative and descriptive points are retained as the points to be registered. Secondly, the unique shape context (USC) description is added to the feature points. Then, coarse matching is completed based on the progressive sample consensus (PROSAC) algorithm. Finally, a bidirectional KD-tree is established for the source and target point clouds to reduce the search time and accelerate the iterative closest point (ICP) to complete the fine registration.