Considering the large amount of calculation, low efficiency, and poor real-time performance of mobile scanning registration when using the traditional point cloud registration method to process large point cloud models, a point cloud registration method based on the convolution neural network combined with the improved Harris-SIFT (Scale Invariant Feature Transform) is proposed. First, the Harris-SIFT algorithm is improved so that it can extract the stable key points of a point cloud model in three-dimensional space. Second, the weighted adjacency matrix of the key points is used as the input feature map for the convolutional neural network. This allows for prediction matching of the key points from the source point cloud and the target point cloud. Then, based on the key points of the matching, the iterative closest point (ICP) algorithm is used to achieve precise registration of the point cloud data. Compared with the traditional point-to-point registration, the proposed method does not need to generate corresponding point descriptors, thus avoiding the problem of a high global search overhead. The experimental results reveal that when compared with the ICP algorithm, the proposed method can better complete real-time point cloud registration, needs minimal calculation, takes a short amount of time, and is highly efficient.