|
||||
|
Telepresence and Teleoperating suffer from a latency of feedback information when carried out on a communication channel with low bandwidth. To compensate for the delayed update information, a local model of the remote scene is maintained to present photorealistic scene prediction. Upon arrival of new scene data, the local scene model is adjusted accordingly. Therefore the operator is always presented with an accurate view of the remote scene, which follows the given commands and movements without waiting for the feedback to update the view. To acquire reliable scene models, a stereoscopic camera system is used to obtain a depth map of the remote scene. A mesh of polygons is laid through the 3D points, which then can be modified in any way by accessing the mesh's nodes. Texture mapping finally provides a photorealistic view of the predicted scene view. In order to calculate the depth map for this scene model, the disparity between pixels in the stereoscopic camera images, correlating to the same scene point, is needed. Due to camera noise, pixel comparison is made by a reference block of neighboring pixels. Using a similarity measure such as normalized cross-correlation |
or sum of absolute differences, a correlation matrix can be calculated for each scanline. This correlation matrix contains the correlation- likelihood values (e.g. match values) of small reference blocks in the left image to small reference blocks in the right image, within a certain search range. All correlation matrices together build the correlation cube. By examining the correlation-likelihood values of this cube, correct matching pixel pairs can be found and the disparities between them are written into the disparity map. The diploma thesis presented here shows a fast way to calculate a correlation cube and describes and evaluates different approaches to obtaining disparity maps from stereoscopic image pairs. In particular, a dynamic programming approach and a cooperative approach are demonstrated and advantages and disadvantages are discussed. The detailed examination of these algorithms leads to new and modified algorithms, which again are evaluated with respect to the needs of teleoperating tasks. A final comparison proposes a strategy on how to improve the most promising algorithm in subsequent work to meet the required real-time demands. |
|||
|
|
||||
