非结构化自然图像自主导航中的两类分割方法

Fast Cut-uncut Segmentation of Textural Natural Image for Autonomous Navigation

以园艺草割过与未割过的自然图像两类分割为研究对象 ,提出了基于多尺度特征提取 ,以行像素最小、次小和最大、次大值求取加权中值点集的窄带兴趣区法 ,以及相邻行像素两类特征证据增强与多证据模糊判别增强分割法。通过自动跟踪分割实验 ,显示出自然图像中的不同色块和阴影对分割影响不大 ,而且能够做到实时输出分割参数 ,70 ms内自主做出未经透视投影变换的作业机理想移动方向决策。这两种方法对纹理两类边缘的分割是完全无监督的 ,可避免耗时的计算和人工操作介入。

Two novel methods are proposed for fast cut-uncut segmentation of textural surfaces in navigating outdoors agricultural robot. The key to the efficiency is based on the narrow band extraction of multi-scale features from interest region. The other crucial factor is based on the multi-evidence enhancement of pixel-rows. The former is related to the weighted mean of highest, sub-higher and the lowest, sub-lower values of pixel-rows. The latter is related to the feature enhancement of neighborhood rows and multi-evidence fuzzy recognition. We can track the boundary between cut-uncut textural surfaces for the autonomous navigation of robotic lawn mower and let it mow in a pattern as a human would. Through local histogram analysis of the cut-uncut grass of natural texture, we found that the two approaches were totally unsupervised and their navigation line was adaptive to changing environments. The software for analyzing the lawn-mowed and non-structured natural images was also developed. The results we obtained are quite promising (correct segmentation was done within 70 ms) and the technique can be further used in real-time tracking and navigating of image sequences.