Point cloud registration technique based on variational coding cross fusion

In light of the fact that the current mainstream deep point cloud registration algorithms merely take into account the extraction of features but disregard the function of spatial positions, this paper puts forward a point cloud registration technique based on variational encoding cross-fusion. The position embedding cross-attention mechanism is used to crossfuse the extracted point cloud features, and enhance the understanding of point cloud features to location information. At the same time, a variational encoder is employed to encode the point cloud features into the latent space, and a variational loss function is utilized to train the model, strengthening the model's ability to distinguish between the overlapping and non-overlapping parts of the point cloud. Experimental results demonstrate that the improved algorithm surpasses the current mainstream point cloud registration algorithms.