关键词: 多镜头全景相机, 球面投影, 球面视频生成, 薄板样条变换

Abstract: Multi-head panoramic camera(MPC) is a kind of integrated optical system with ultra-wide field of view and can image an entire hemisphere onto a flat image plane through image stitching technique. But for actual scenes, the spherical video outputted by MPC often generates undesirable visual artifacts. This paper presents a method to generate seamless spherical video suitable for any scene. It mainly includes two aspects:① Estimation of MPC spherical projection parameters. Taking the tie points(pixels) detected in overlapping area as the observation values, those parameters are automatically estimated though minimizing the angle between the spherical projection center and corresponding spherical space points with the least square technique. As a result, the MPC video stitching errors caused by inaccurate calibration parameters, depth change of scene content and the inconsistency between MPC are effectively reduced; ② Establishment of TPS model for MPC spherical video generation. This model is established using the spherical re-projection geometry of MPC video pixels as global transformation and estimated with the salient pixels along selected seam line. As a result, video pixels of MPC sub-cameras can be directly mapped to the stitched spherical video with minimal pixel error in the overlapping area and the visual artifacts are completely eliminated simply by linearly mixing the pixels around selected seam line. The experimental results on stimulated and real scene show that, proposed method can generate seamless spherical Panoramic video for MPC used in various scene. In addition, proposed method is implemented efficient and can meet the requirements of high frame rate video output of MPC, and has good application value.

Key words: multi-head panoramic camera, spherical projection, spherical video generation, thin plate spline