Curvature Estimation And Data Simplification Of Point Clouds Based On Improved PointNet

Point clouds is an important contents of computer vision,and the key in such hot fields as autonomous driving and digital museums.On the one hand,most of the traditional algorithms based on non-deep learning have problems such as manual design features and vulnerability to noise points.On the other hand,PointNet based on deep learning shows its advantages in point clouds segmentation,classification and other applications,such as the ability to adapt to variable data sets and strong anti-noise ability,but it cannot to estimate curvature of point clouds or to simplify data directly.In view of the above situation,in this paper,we propose an algorithm of point clouds that can estimate curvature and simplify data based on improved PointNet.Firstly,we propose a network that can estimate curvature of point clouds based on improved PointNet,which consists of sampling layer,feature extraction layer and regression layer.The sampling layer conducts local point clouds through KD-Tree.The feature extraction layer with PointNet as the main structure extracts the local feature of each local point clouds obtained by the sampling layer.By using quaternions to replace the 3×3 affine transformation matrix in PointNet,the point is only transformed by rotation,which solves the problem of unstable curvature estimation in PointNet due to scale change.The regression layer uses multi-layer perceptron to predict the curvature of each point.Then,we design a weighted RMSE as loss function.By using the relative error of high curvature points and the absolute error of low curvature points,the error values of different curvature points have similar weight in calculation.The Bunny and other 3D models in The Stanford 3D Scanning Repository were used to verify our algorithm,and the absolute error of mean curvature as evaluation criterion.The results show that for the noisy point clouds,our algorithm is better than the pseudo inverse quadratic fitting and Taubin quadratic fitting,and the MAE can be reduced by 60%.Secondly,a data simplification algorithm based on point clouds curvature estimation network is proposed.Firstly,in the curvature estimation network,the curvature features extracted of points from the feature extraction layer are mapped by one-dimensional maximum pooling to obtain the original point clouds.Then,we weight the points according to the number of times they are mapped.Finally,the point clouds are simplified according to the weight.For the N × 3 points,the simplified 3D model can be output directly.The results show that the algorithm in this paper can effectively deal with 3D models with large curvature changes and complex shapes.It can not only retain the geometric features and contours of the original 3D model,but also has a low simplification error.Compared with the 3 algorithms,such as clustering simplification and fast resampling,the average simplification error can be reduced by 6% at a simplification rate of 50%.Through experiments on the curvature estimation and data simplification algorithm respectively,the feasibility and effectiveness of the curvature estimation and data simplification of point clouds based on improved PointNet are verified.