无人车行驶环境特征分类方法
Classification method of running environment features for unmanned vehicle

为了提高车载2D激光雷达对城市环境障碍物的分类能力与环境地图创建精度和无人车自主行为决策的安全性与准确性,提出了一种基于机器学习的环境特征分类方法。将2D激光雷达的观测数据帧分割为独立的数据段,每个数据段中包含一个环境障碍实体; 在数据段的二维高斯概率密度空间中,以概率密度的等高线椭圆轴长、对数似然值和最大概率密度作为人工神经网络的样本数据元素,利用人工神经网络完成数据段分类; 利用人工神经网络输出值的权重对分类的有效性进行判定,仅保留有效的环境特征,并对分类完成的观测数据进行特征提取。计算结果表明:在同一个试验场景中,当分类有效性判定条件被设定为分类稳定区间为[0.55,1],分类过渡区间为[0.45,0.55),分类无效区间为[0,0.45)的宽松条件时,共识别出98个环境特征,同一环境特征的多次观测数据的分类提取结果之间的最大标准差为30.7 mm,多个环境特征的平均标准差为5.1 mm; 当分类有效性判定条件设定为分类稳定区间为[0.65,1],分类过渡区间为[0.35,0.65),分类无效区间为[0,0.35)的严格条件时,共识别出93个环境特征,同一环境特征的多次观测数据的分类提取结果之间的最大标准差为22.0 mm,多个环境特征的平均标准差为4.2 mm,因此,提出的分类方法的噪声容忍能力强,分类精度高。
In order to improve the barrier classification ability of mobile 2D LiDAR in urban environment, the creating accuracy of environmental map, and the safety and accuracy of autonomic behavior decision-making for unmanned vehicle, a classification method of environmental features based on machine learning was proposed. The data from 2D LiDAR were divided into independent data segments, and each data segment contains one environmental barrier. In 2D Gaussian probability density space of data segments, the elliptical axial lengths of contour lines, the log likelihood values and the maximum density were taken as the elements of sample data of artificial neural network, and the data segments were classified by the artificial neural network. The classification validity was estimated according to the weights of artificial neural network’s output data to retain the effective environmental features, and the features were extracted from the classified data. Computational result shows that in the same test scenario, when the judging condition of classification validity is relaxed, under which the classification stability interval is [0.55, 1], the classification transition interval is [0.45, 0.55), and the classification invalid interval is [0, 0.45), 98 environmental features are extracted, the maximum standard deviation of classified extraction results for the multiple observation data of one environmental feature is 30.7 mm, and the average standard deviation for all features is 5.1 mm; when the judging condition of classification validity is strict, under which the classification stability interval is [0.65, 1], the classification transition interval is [0.35, 0.65), and the classification invalid interval is [0, 0.35), 93 environmental features are extracted, the maximum standard deviation of classified extraction results for multiple observation data of one environmental feature is 22.0 mm, and the average standard deviation for all features is 4.2 mm. Therefore, the proposed classification