Self-organization of environmental maps based on scene images and navigation of mobile robots

1) otsu提出的局部自相关函数适用于压缩场景图像。以最小位置估计误差为准则，明确了1/64的特殊分辨率是最佳的。2)在自组织构建环境地图时，通过将位置信息添加到每个场景图像的局部自相关函数中来输入SOM。机器人的位置信息受到噪声的破坏。与传统的贝叶斯估计相比，先验概率分布假设为高斯分布。通过最小化位置估计误差来确定位置信息的相对权重。3)真实的机器人在包含瓦楞纸板障碍物的场地中进行测试。通过实验测量，机器人在前进过程中，由于车轮的滑动而导致定位和方向误差增大。如上所述，包含误差的位置信息被添加到局部自相关函数中，并用作som的扩展输入。4)当给定新的输入向量时，贝叶斯估计可以有效地使用构建的环境图来估计位置。与传统的贝叶斯估计相反，某一时刻的先验概率是由于机器人的运动而导致的前一时刻的后验概率的转换。5)由于环境地图不包含障碍物的位置，很难将位置估计和避障结合起来。我们提出用概率分布的形式来表示红外传感器与障碍物之间的距离，并通过仿真实验验证了其有效性。

1) Local autocorrelation functions proposed by otsu is appropriate for compressing scene images. We clarified that special resolution of 1/64 is the best based on the criterion of minimizing location estimation error.2) In constructing environmental maps by self-organization, inputs to SOM are made by adding location information to local autocorrelation functions of each scene image. Location information of a robot is corrupted by noise. In contrast to conventional Bayesian estimation, prior probability distribution is assumed to be Gaussian. A relative weight of location information is determined by minimizing location estimation error.3) Real robot is examined in a field including obstacles made of corrugated cardboard. As the robot proceeds, location and direction errors increase due to slipping of wheels, which is experimentally measured. As mentioned above, location information including error is added to local autocorrelation functions and is used as an extended input to SOM.4) When a new input vector is given, Bayesian estimation is effective in estimating location using constructed environmental maps. In contrast to conventional Bayesian estimation, prior probability at some time is a translation of a posterior probability at previous time due to movement of a robot.5) Since an environmental map does not include the location of obstacles, it is difficult to combine location estimation and obstacle avoidance. We propose to represent the distance between an infrared sensor and the an obstacle in the form of probability distribution, and demonstrate its effectiveness by simulation experiments.