面向盲人避障的场景自适应分割及障碍物检测

Adaptive Scene Segmentation and Obstacle Detection for the Blind

针对盲人避障系统的实时性和鲁棒性要求,提出了基于单帧RGB-D图像的场景自适应分割和障碍物检测方法.针对室内场景存在较多的平面结构的特点,首先利用深度信息进行由粗到细的两阶段平面快速提取;然后提出了基于图论的RGB-D场景自适应分割算法,将基于深度信息的平面分割结果和基于RGB的图像分割进行自适应的权重融合;最后提出了多层次的障碍物检测识别策略,将场景区域识别为地面、桌面、墙面以及其他障碍物,并给出了场景分析结果的简化模型.实验结果表明,文中方法结合深度数据的优势以及场景的几何结构特点,有效地提高了场景分割和障碍物检测的实时性和鲁棒性,能够满足盲人室内避障应用需求.

To develop real time and robust travel aids system for the blind, this paper introduces an adaptive scene segmentation and obstacle detection algorithm based on RGB-D data. Firstly, a rough-to-fine plane segmentation algorithm based on depth data is used, which makes use of the indoor structure information. Secondly, a graph-based scene segmentation algorithm is introduced, which adaptively combine the results of plane segmentation and RGB image segmentation. Thirdly, a multi- level object detection strategy is applied to recognize segmentations as ground, desk, wall or other obstacle. Finally, a simplified 3D reconstruction of the scene is used to demonstrate the analysis results. The experimental results show the proposed method effectively improves performance of scene segmentation and obstacle detection, which well combines depth data and geometry structural information of indoor scene. The proposed algorithm is fast and robust, which can be applied to indoor obstacle avoidance system for the blind.