二维激光雷达数据角点特征的提取

为增强无人车识别行驶环境中角点特征的鲁棒性, 并提高角点特征的识别速度, 基于观测点的二变量正态概率密度映射之间的相对差值, 提出了一种角点特征提取方法; 将观测数据组映射到二变量正态概率密度空间, 获得每个观测点的映射; 对映射结果进行归一化, 消除协方差引起的数值差异; 在映射数值曲线中寻找波峰与波谷的位置, 波峰对应的观测点最接近均值点, 波谷对应的观测点最接近拐点; 利用波峰和波谷的相对高度判定该组观测数据是否符合角点特征的边长要求; 用波谷对应的原始观测数据点坐标作为角点特征, 构建环境特征地图。试验结果表明: 提取方法能够处理观测点数大于63, 观测点角度分辨率大于1°的稀疏观测数据, 在大尺寸室外环境和室内环境中, 提取方法都能够稳定识别大型角点; 对小于180个点的观测数据, 最大处理时间小于5ms, 平均处理时间小于1.9ms, 提取方法减少了构建环境特征地图的时间; 提取方法依据观测数据的二变量正态概率密度提取角点特征, 对观测误差和角点特征的尺度与形状不敏感, 能够有效提高角点特征的识别鲁棒性。 

Corner feature extraction of 2D lidar data

In order to enhance the robustness of the corner feature recognition in the driving environment by the unmanned vehicle and improve the recognition speed of the corner feature, based on the relative difference between bivariate normal probability density map values of observation points, a corner feature extraction method was proposed. The observation data set was mapped to the bivariate normal probability density space, and the mapping value of each observation point was obtained. The mapping results were normalized, and the numerical differences caused by the covariances were eliminated. The positions of peaks and troughs were found in the mapped numerical curve. The observation point corresponding to the peak was closest to the mean point, and the observation point corresponding to the trough was closest to the inflection point. Whether the set of observed data meets the edge length requirement of the corner features was determined by using the relative heights of peaks and troughs. Thecoordinates of the original observation data points corresponding to the troughs were used as corner features to construct the environment feature map. Test result shows that the extraction method can process sparse observation data with more than 63 observation points and angular resolution of the observation point greater than 1°. Therefore, in large-scale outdoor environment and indoor environment, the extraction method can stably identify large corner points. When the observation data points are less than 180, the maximum processing time is less than 5 ms, and the average processing time is less than 1.9 ms, so the extraction method has good real-time performance, which is conducive for decreasing the time required for designing the environment feature map. The extraction method extracts the corner features according to the bivariate normal probability density of the observation data, is insensitive to the observation error and the scale and shape of the corner feature, and can effectively improve the robustness of corner feature recognition.14 figs, 25 refs.