The Research On 3D Reconstruction Of Objects And Scenes Based On The Depth Camera

In the era of artificial intelligence,3D reconstruction is an important technology for intelligent devices to sense and interact with the surrounding environment.It is widely used in robot vision,driverless and other fields.There are still many problems in the application of 3D reconstruction for non-rigid bodies and large scenes.For example,due to the soft texture of a non-rigid body,it is easy to undergo flexible deformation during images acquisition,which leads to mismatch and trajectory drift.Large-scale scenes cover a wide area,so the camera used to collect data has long trajectories,and the pose error will continue to accumulate during the reconstruction process,which easily causes the camera to lose tracking.The basic idea of the reconstruction research on non-rigid body and large scene is to use the depth data to realize the three-dimensional reconstruction by dividing the appropriate data processing interval,merging the interval data,and reducing the interval data fusion error.Aiming at the problem of mismatch and trajectory drift caused by non-rigid deformation during the data acquisition phase,this paper proposes an adaptive segmentation strategy based on the cumulative error deduced from the number of effective corresponding points.It divides the camera pose estimation into multiple data processing stages independently to overcome the continuous accumulation problem of the pose estimation error,then forms a complete camera motion trajectory.Aiming at the unsmooth surface between the different data processing stages,namely the crease problem,this paper combines rigid optimization and non-rigid optimization.In the rigid optimization,based on the ICP algorithm,the model-tomodel energy function is proposed to optimize the rigid transformation matrixes in order to reduce the crease.In the non-rigid optimization,the non-rigid constraint based on the affine transformation is added to make sure the final non-rigid 3D model is complete and smooth.In order to improve the detail of the 3D model of large scenes,this paper proposes a scene subvolume reconstruction algorithm combining depth and color data.Color data provides texture information when geometric information is missing,making camera pose estimation more accurate.In order to prevent the pose estimation error from accumulating during the reconstruction process,the subvolume is moved along the camera trajectory.The pose tracking and data fusion are performed separately in each subvolume,and the subvolume moving strategy is proposed for the case of large rotation of the camera.The LM method is used to optimize the poses of the subvolume in the global coordinate system,and further reduces the pose estimation error by setting the step coefficient and optimizing weight.In this paper,the proposed three-dimensional reconstruction algorithm of nonrigid body is verified by human body three-dimensional reconstruction experiments.The large-scale three-dimensional reconstruction algorithm proposed in this paper is verified by self-extraction data and public data set experiments.