Research On Probability Driven Approach For Point Cloud Registration Of Indoor Scene

Many conveniences has been caught to people by the development of science and technology.The 3D data forms emerge endlessly,and gradually change people's life style and thinking mode.Point cloud can represent the space shape and color information of objects or scenes,and is often used in the construction of 3D models.3D point cloud data can be acquired by 3D scanners or RGB-D cameras.But due to the camera's limited view angle,the current 3D scanning equipment can only obtain the point cloud data within a certain angle of view at certain time.Therefore,the point cloud registration algorithm is needed to align the point cloud data from different angles.In computer vision field,3D point cloud registration is also a classic topic for many years.It is of great significance in the fields of 3D modeling,animation design,cultural relics restoration,medical model processing and so on.The existing point cloud registration algorithms are usually divided into two stages:coarse registration and fine registration.The former roughly aligns the point clouds,the latter focuses on the difference of details between the point clouds,and aligns the detail parts between the point clouds.These methods can obtain considerable registration results on the data set of better quality.However,for the general point clouds gained by low resolution camera,or contains many noises,and some indoor scene point clouds with many occlusion or missing information,and without obvious geometrical properties.These algorithms cannot get ideal registration results through relatively complex calculation process.Aiming at the problem of how to achieve the ideal registration result of low-resolution indoor scene point cloud,this paper explores and studies the registration related knowledge,and proposes a probability driven approach for point cloud registration of indoor scene.The main works are as follows:(1)In this paper,the distance matrix and the difference matrix that based on scalarization and vectorization are constructed separately,which are used to represent the coordinate and position relationship of key points,and act on the calculation process of corresponding probability of key points;(2)In this paper,a probability-based method is proposed to calculate the matching probability between the key points of two point clouds,and the idea of distance matrix and difference matrix is combined with probability algorithm to realize the indoor scene point cloud registration.(3)Aiming at the few failures of the above algorithms,this paper proposes a registration optimization algorithm combining color information and quaternion.In this method,color information is combined into the registration algorithm,and the complex and traditional matrix form is replaced by a simple quaternion representation.In the experiment,the paper conducts a lot of ablation experiments and comparative experiments based on several RGB-D data sets and different specifications of indoor scenes,to certify the availability of this algorithm.The results of the experiment display that the method can obtain better registration effects and faster processing speed on this data sets,especially for the scene point cloud with low resolution,complex scene and no prominent features.