Robust Monocular Pose Tracking of Less-Distinct Objects Using Contour Part Model
BI DaoMing,HUANG Hui,LIU JinBo,SUN XiaoLiang,YU QiFeng
DOI: https://doi.org/10.1360/sst-2019-0336
2020-01-01
Scientia Sinica Technologica
Abstract:Less-distinct objects are those that are similar to the background in terms of color or texture. The presence of interference, e.g., noise, dynamic light, and cluttered background, further reduce their distinctiveness. Owing to the lack of distinctiveness with the background, robust monocular 6 DOF pose tracking of less-distinct objects remains an unresolved issue. We propose a novel contour part model based on a robust monocular pose tracking method for less-distinct objects. This work uses the traditional contour feature for monocular pose tracking in a new strategy called a contour part model. First, the contour part model is built segmenting the projected contour rendered from the 3-D model into contour segments of a certain length adaptively, according to the Shi-Tomasi cornerness scores. Further, for each contour part, the correspondence is detected in the input image by the proposed gradient orientation-based template matching. Finally, robust pose tracking is achieved solving the Perspective-n-Point problem. Results of the experiments carried out using a publicly available dataset and semi-synthetic images indicate that the proposed method performs better than the current methods when pose tracking less-distinct objects, showing great robustness toward interference. The superior performance of the proposed method is validated by the experiments.