以圖切分演算法搭配物理支持關係在擁擠都市中完成移動式RGB-D相機之運動分割

Abstract

在計算機視覺與機器人學領域運動分割是一個重要且具有挑戰性的問題，由於其複雜性，根據環境差異性、運動物體本身的特性和所運用的感測器，這個問題可以從不同研究方向切入。透過使用 RGB-D 結構光距離感測器，我們提出一個新穎且強健的算法來完成運動分割。我們利用色彩和空間上的一致性改進隨機抽樣一致算法（RANSAC）來進行運動估測，它繼承了原有 RANSAC 範式下的計算效率與強健性，並且我們結合基於運動估測結果的內群和離群分佈和 RGB-D 影像上的色彩與深度資訊，以圖切分演算法完成運動分割，進一步，我們引入物理支持關係除了用來表示運動分割強度，並且可以用來理解所分割運動之區段的語義。相較於現有類似架構之運動分割算法，此算法在一個高度動態的環境有良好的表現。最後，我們提供了一個在擁擠都市環境下使用 Xtion Pro Livee 感測器所錄製的 RGB-D 數據集來展示我們成果。

Motion segmentation is an important and challenging problem in computer vision and robotics. Because of its complexity, this problemcan be approached from respective angles depending on sensors, environment, andmotions themselves. By utilizing RGB-Dvideo captured by a structured light range sensor, we proposed a novel and robust algorithm to segment motions from consecutive frames. Based on a modified random sample consensus algorithm (RANSAC), we exploit the coherence of color and spatiality in the scene to estimate motion. It inherits the computational efficiency and probabilistic robustness from the RANSAC paradigm. After aligning two point clouds by the estimated transformation, we combine the output of inlier and outlier distribution with the prior knowledge of the RGB-D images to conduct segmentation by a graph-cut optimization scenario. Moreover, we introduce physical support relationships to better understand the motions in the environment. We provide a RGB-D dataset captured in a crowded urban environment to demonstrate our idea. Comparing to several motion segmentation methods in the same pipeline, we show that our approach performs well in the highly dynamic scene.