Fig. 1. Pipeline of proposed method

Fig. 2. Schematic diagram of normal vector ambiguity

Fig. 3. Figure （a） shows that when the point （marked red） to be judged is outside the model， the intersection of the yellow ray and the model （marked green） has two even numbers. The point （marked red） to be judged in the figure （b） is inside the model， and there is an odd number at the intersection of the yellow ray and the model （marked green）

Fig. 4. （a） shows the method of obtaining the plane perpendicular to qin， and pi+1 is the nearest sampling point from pi， in （b）， the skeleton point qi is in an unbalanced position， and in （c）， the skeleton point qi is in an balanced position

Fig. 5. Difference between R and RL. （a） is R， and （b） is RLobtained through breadth first search according to R

Fig. 6. Curve skeleton of some models （right） and the display of the original model （left） extracted by this algorithm

Fig. 7. Effect comparison of curve skeleton extracted by various methods

Fig. 8. Effect comparison of curve skeleton extracted by various methods when inputting the model with noise

Fig. 9. Action transformation of lidar point cloud based on the guidance of curve skeleton

Table 1. Statistics of the number of skeleton points， topological connections， junction points and topological connection errors of several methods

Table 2. Running time of different algorithms