全景摄像机

全景相机数据集

• LIBOMNICAL: [一个标定工具，也提供了折反式相机数据与全景相机数据] provides a MATLAB Toolbox to calibrate central and slightly non-central catadioptric cameras and catadioptric stereo setups. This is a reference implementation of the centered projection model presented in Calibrating and Centering Quasi-Central Catadioptric Cameras (ICRA 2014) as well as a number of central reference models. The calibration consists of an automatic corner extraction stage, a minimization part and the visualization of the calibration results. The toolbox contains demo images to run a quick demo calibration of a slightly non-central catadioptric stereo camera setup, where the corners of the images are already extracted. After installation of the library the calibration runs using a single command as described in the readme file. All results are saved to a calibration folder as illustrated below. Besides the calibration result, the folder will contain the calibration images with the detected and estimated checkerboard positions (left), a visualization of the reprojection error for all checkerboards of each camera (middle) and the reconstructed 3D positions of the calibration patterns (right).
  

• OCamCalib: [全方位相机标定工具，提供了Matlab的实现] Omnidirectional Camera Calibration Toolbox for Matlab (for Windows, MacOS & Linux)
  
  The OcamCalib Toolbox for Matlab allows the user (also inexpert users) to calibrate any central omnidirectional camera, that is, any panoramic camera having a single effective viewpoint (see section 17). The Toolbox implements the procedure initially described in the paper 1 and later extended in 2 and 3. A detailed introduction to this model is in section 19 of this Tutorial. Furthermore, you can also see a demo of how the toolbox works here.
  The Toolbox permits the user to easily and quickly calibrate the omnidirectional camera through two steps. First, it requires the user collect a few pictures of a checkerboard shown at different positions and orientations. Then, the user is asked to extract the corner points. With the new version of the toolbox this operation is done completely automatically. Therefore, no manual extraction is needed. After these two steps, the calibration is completely automatically performed.

After the calibration, the toolbox provides two functions (CAM2WORLD and WORLD2CAM) which express the relation between a given pixel point and its projection onto the unit sphere (this is a 3D vector emanating from the single effective view point) (see section 17). This relation clearly depends on the mirror shape and on the intrinsic parameters of the camera. The novel aspects of the OCamCalib Toolbox with respect to other toolboxes are the following:
The toolbox is the only one with Automatic Corner Extraction (no manual extraction is required).
The toolbox does not require a priori knowledge about the mirror shape.
It does not require calibrating the perspective camera separately: the system camera-mirror is treated as a unique compact system that encapsulates both the intrinsic parameters of the camera and the parameters of the mirror.
The detection of the image center is performed automatically. It does not require the visibility of the circular external boundary of the mirror. Unlike other toolboxess, which require the visibility of the external boundary of the mirror to determine the image center, the OCamCalib Toolbox automatically identifies the center without any user interaction!
The calibration performed by the OCamCalib Toolbox is based on the following hypotheses:
The camera-mirror system possesses a single effective viewpoint (see section 18 for a definition), or also a “quasi” single viewpoint. In fact, the Toolbox is able to provide an optimal solution even when the “single view point property” is not perfectly verified (for instance when the camera optical center is not exactly in the focus of the hyperbola or also for spherical mirrors). The Toolbox showed to give very good calibration results even in the latter case (reprojection error < 0.5 pixels!).

• Yalin: 分类、人、车的检测
• ImageCLEF:

应用中存在的问题

• 最主要的原因是全方位摄像机监视场景大，但输出图像分辨率与普通摄像机相同，因而在当前图像分辨率下，虽然可以覆盖多个普通摄像机的监视范围，但不能提供场景中特定目标的清晰观察（例如人脸），因此，在很多视觉监视场合难以应用。

• 严重变形，相对低且不均匀的分辨率，目标在全方位图像中成像面积小导致检测的灵敏度降低，目标的可视性差。

• 解决方案：全景相机+枪机联动 提高观测质量，（能否超分辨率重构？）