| 研究実績の概要 |
There are two main results from this year.
1) Additional experiments revealed that shadows cast on the object seem to be a bigger problem than occlusions, which was discussed in last year’s plan. The method was not able to adapt quickly enough, and deemed the shaded part on the object to be background. A separate process to classify shadows has been developed and its results are dynamically incorporated into the object model to circumvent this problem. The accuracy has also been improved by dynamically adapting the resolution of the contour identifying the object to better capture uneven parts of the boundary resulting in a contour that is more true to the object shape. Finally the algorithm has entirely been moved to the GPU for significantly better performance, allowing for processing of > 500 Hz HD video. Currently a conference paper and a journal paper are being prepared based on these results.
2) As proposed last year, a feature point based tracking method has been investigated, specifically targeting rigid objects. It exploits the fact that the distances between points on the object are the same when the object moves or rotates. By tracking these points as the object moves, points that diverge from these distances can be discarded. By assuming the object can be (partially) contained in a cuboid, new points can be added to the object model as new sides of the object become visible. It was shown that in addition to tracking the object, its pose could also be recovered. This method has been developed using both 2D and 3D image data.
|
