Improved range image segmentation by analyzing surface fit patterns

We propose a new approach to range image segmentation of planar and curved surface scenes. Our method is mainly an extended design of an existing algorithm, which was guided by a framework of performance evaluation. We choose the range segmentation algorithm developed by Jiang and Bunke as our baseline algorithm, which is fast and has shown relatively high performance in several experimental performance evaluation studies. We analyze the types of errors made by the algorithm, propose design modifications to decrease the error rate, and experimentally verify that the new approach achieves statistically significant performance improvement. Whereas the baseline algorithm applies the edge-linking uniformly to all edge pixels to segment a region, the modified algorithm selects high potential edge areas in the region by analyzing the surface fit pattern and gives priority of edge-linking to those areas. The contributions of this work are (1) an improved algorithm for segmentation of range images of both planar and curved surface scenes, and (2) a demonstration of using empirical performance evaluation to guide algorithm design and modification to achieve better performance.