3D Point Cloud Scene Understanding Based On Bayesian Methods

With the technological development of 3D sensors,many computer vision tasks related to3 D point clouds emerge as the times require,among which 3D scene understanding tasks face many challenges.In the past few years,the use of deep neural networks for 3D object detection and scene classification on point clouds has gradually become the mainstream method.However,traditional neural networks only use shape features to identify objects and scenes,and do not consider the association between objects and scenes.In this thesis,the Bayesian neural network method is used to endow 3D object detection with the ability to predict uncertainty.The Bayesian belief network is used to infer about the scene,and the rationality of the object detection result is verified,thereby reducing unreasonable outputs in 3D object detection,and adding interpretability for scene classification.Aiming at the problem that the traditional 3D object detection model cannot evaluate the uncertainty,this thesis improves the Group-Free 3D object detection model according to the principle of Bayesian neural network,and proposes the Bayes GF3 D model,which makes the point cloud object detection model able to predict uncertainty.In addition,this thesis proposes the concept of Bayesian convolutional layer and improves the calculation process of Bayesian fully connected layer.In experiments,the Bayes GF3 D model can effectively predict the uncertainty of objects without significantly reducing accuracy.The thesis also discusses the influencing factors of the uncertainty of object detection in point clouds.Aiming at the problem of using object categories to reason about the scene,a probabilistic graph model of 3D point cloud indoor scene based on Bayesian belief network is proposed.Classes of indoor scenes can be efficiently predicted from observed object evidence.Aiming at the problem that the traditional neural network cannot make full use of the context information of the scene,a scene understanding model based on the relationship between objects and scenes is proposed.The results are verified for rationality,thereby improving the accuracy of 3D point cloud object detection and obtaining an interpretable scene understanding model.The proposed model for point cloud indoor scene understanding can effectively filter out the detection results that do not conform to the environmental context in 3D object detection,thereby improving the accuracy of object detection,and proposes an interpretable model for the task of scene classification.