无人车越野环境感知关键技术研究

Involving computer image processing, pattern recognition, multi-sensors data fusion and multidisciplinary theory, this project explore new way of environmental perception based on analogue and computation of cognitive mechanism. The key technology of autonomous navigation for unmanned ground vehicle (UGV) working in field environment would be developed to implement a cross country UGV environmental perception technical framework which enables information interact among sensors. A way based on probability test and Gaussian Mixture Model to obtain the running region of the UGV are proposed in this subject, which applies methods by analyzing the principal component such as distance contrast of three-dimension data, image edge chain-code curvature, covariance matrix and so on. Characteristic vectors of the objects are gathered from each sensor. Then the subordination obtained by using the fuzzy interpolation is applied to calculate the basic probability assignment. It is supposed that the subordination is equal to correlation coefficient in the formula. More accurate results of object identification would be achieved by using the D-S theory of evidence. Control on motion behavior of cross-country UGV and autonomous navigation are based on this theory, which is a necessary pre-condition for realizing UGV high speed driving in cross-country environment.

本项目借鉴计算机图像处理、模式识别、传感器信息融合等多学科理论知识，在认知机理模拟和计算的层次上探索环境感知的新方法，研究面向越野环境的无人车自主导航关键技术，初步实现具有多传感器信息交互的越野无人车环境感知技术框架。本课题提出基于概率检验法和混合高斯模型探测环境中可行驶区域，并利用三维数据的穿透率、图像边缘链码曲率、协方差矩阵的主成分分析等方法，提取观测目标来自每个传感器的特征向量，采用摸出插值法确定隶属度以及相关系数构造基本概率赋值函数，基于D-S证据融合理论产生更精确的目标身份分类识别结果，它是越野无人车实现运动行为控制、自主导航的分析依据，也是实现野外环境高速行驶的必要前提条件。

地面无人车辆在野外环境感知会碰到多种正、负障碍物，如草丛、岩石、斜坡、树木、水域等等，本项目借鉴计算机图像处理、模式识别、传感器信息融合等多学科理论知识，研究具有多传感器信息交互的越野无人车环境感知技术框架，开展以下主要研究工作：（1）本项目基于一阶Markov模型纠正激光系统误差，选用“概率检验法”探测车体前方障碍物，并选择判别型学习算法“坐标上升法”局部优化阈值参数及马尔科夫链的误差参数。（2）基于激光与视觉图像的投影映射提取视觉图像中近距离可行驶表面，通过训练该表面的混合高斯模型GMM将视觉图像分类成可行驶区域和障碍物区域。（3）基于单目视觉与激光三维数据信息，利用图像边缘链码曲率、激光点云的穿透率、协方差矩阵的主成分分析等方法，提取观测目标来自每个传感器的特征向量。（4）基于马尔可夫随机场模型与Bayes理论方法分类识别障碍物，结合GPS、微型陀螺仪建立UGV当前环境感知区域地图并确定可行驶区域。