Robust Model Based Motion Segmentation

This paper presents a new algorithm for the motion segmentation task. The proposed algorithm addresses the important issue of the interconnectivity behveen data segmentation, model selection and noise scale estimation. The algorithm is tested on motion segmentation of multiple objects undergoing different types of motion. The results of applying our algorithm to range data segmentation are also included. 1. I n t r o d u c t i o n A broad spectrum ofcomputer vision applications depend upon accurate and’ reliable data segmentation algorithms. In particular, motion segmentation is an important part of many robotic applications: such as tracking and path planning. A novel approach to visual data segmentation i s presented in this paper. The proposed method deals with the two crucial issues of data segmentation: model selection and noise scale estimation, simultaneously. Generally speaking, the proposed algorithm can be categorised as a parametric (model based) technique although some of i ts steps make use o f spatial information (for sampling and model selection). The rest o f this paper is organised as follows. We briefly review some o f the visual data segmentation literature in section 3. As our proposed method relies on model selection criteria, we dedicate section 2 to the analysis o f different model selection criteria in motion segmentation context. A brief overview o f motion and range segmentation literature is provided in sections 3 and 4. The procedure for estimating the scale o f the noise is described in section 5. In section 6, the proposed segmentation algorithm ;is explained and the results o f our experiments on motion and range data segmentation are presented in section 7. Section 8 concludes the paper. It should he mentioned that the segmentation algorithm described in section 6 is an improvement o f our previous work [2] by adding some stages to incorporate a model selection criteria in the segmentation process. The paper also includes a survey o f different model selection criterions. 2. Model Selection In all the parametric image segmentation algorithms, one needs to find the most appropriate model which would hest fit the data. Therefore, a suitable model selection criterion should he viewed as an important part o f any parametric segmentation algorithm. We have investigated a wide range o f model selection methods including: AIC, G-AIC, CAIC, CAICF, CP, SSD, BIC, MDL, G-MDL, GBIC, BMSC-BAYES for motion segmentation and we have concluded that G-MDL (proposed by Kanatani [7]) performs better than any o f the above techniques (see [3] for more details). 2.1. Test ing D i f f e ren t M o d e l Selection Tech-