Three-dimensional point cloud visual tracking method and system based on multistage feature fusion

The invention discloses a three-dimensional point cloud visual tracking method and system based on multistage feature fusion, and belongs to the technical field of three-dimensional point cloud tracking. Comprising the following steps: acquiring a template frame and a search frame, executing a feature extraction operation combining point attention embedding on the template frame and the search frame in parallel, and respectively acquiring point features corresponding to different levels; mutually fusing the point features corresponding to different levels according to a residual network mode by using cross attention to obtain a fused depth feature; performing rough prediction according to the fused depth features to obtain a rough prediction result; and performing refined prediction according to the rough prediction result, the final output of the feature extraction operation and the fused depth features to obtain a refined prediction result. According to the method, full extraction of point cloud features ca