Vectorization of floor plans (VFP) is an object detection task that involves the localization and recognition of different structural primitives in 2D floor plans(FP). The output of VFP can be further processed for the purpose of plan reconstruction, 3D reconstruction and automatic furniture layout. So far how to make the existing 2D floor plans vectorized faces the problem of recognition inaccuracy and inefficiency. This paper proposed a floor plan recognition algorithm based on key points, which is meaningful and useful. First, the algorithm identifies the effective subject of the FP with the help of the object detection algorithm; then, it builds a deep backbone network to identify the key points and semantic information of the marked plane elements; finally, the algorithm utilizes the post-processing algorithm to optimize and retrieve vectorized data information. Compared with existing methods, the algorithm adopted in this paper enhances the support for the recognition of elements such as sloping walls and bay windows, and effectively improves the recognition accuracy.