No abstract

No abstract

An intelligent robotic living assistive system has become a popular research in the last decade. One of the important topics in that research area is 3D object reconstruction from multiple views. This process may depend on motion estimation using vision. However, often a domestic robot on an electric wheel chair has to move in a steep rotational angle that causes motion estimation from vision to become inaccurate. In addition, an oblique viewing angle creates a perspective distortion to the captured images, which further worsens the estimation result. Hence, in this paper, we propose a new approach by altering the motion estimation problem into a 2D image registration problem. Our method's accuracy is very close to that of the Scale Invariant Feature Transform (SIFT) features tracker, whereas the Kanade-Lucas-Tomasi (KLT) tracker's drops as soon as the rotational angle reaches about 40 ∘ . Although our method is 2.7 times slower than the KLT tracker, it is 19 times faster than the SIFT tracker.

Segmenting a scene into meaningful regions is a classical problem. One of the methods is segmentation from Stereo (on condition that the object has sufficient texture). Most of the time, our interest is foreground objects. However, in dense stereo matching the entire image is processed, thereby wasting time. Hence, in order to process only the foreground objects, we insert a preliminary process to discard the background using dominant motion estimation. Our method estimates the background from stereo images using the phase correlation method. The dominant background area, which has high phase correlation, is then eliminated leaving the foreground to be processed by the traditional Sum of Squared Differences (SSD) algorithm. Thereby, the time taken for the overall process is reduced.

No abstract