一种深度学习辅助的探地雷达定位方法

在恶劣条件下，例如在雨、雪、沙尘、强光以及黑夜等环境下，自动驾驶方案中常用的视觉和激光传感器因为无法准确感知外界环境而面临失效问题。因此，该文提出一种通过深度学习辅助的探地雷达感知地下目标特征用于车辆定位的方法。所提方法分为离线建图和在线定位两个阶段。在离线建图阶段，首先利用探地雷达采集地下目标的回波数据，然后使用深度卷积神经网络(DCNN)提取采集的地下回波数据中的目标特征，同时存储提取的目标特征和当前地理位置信息，形成地下目标特征指纹地图。在定位阶段，首先利用DCNN提取探地雷达采集到的当前地下回波数据中的目标特征。然后基于粒子群优化方法搜索特征指纹地图中与当前提取的目标特征最相似的特征，并输出该特征的地理位置信息，作为探地雷达定位车辆的结果。最后利用卡尔曼滤波融合探地雷达定位结果和测距轮测量的里程信息，得到高精度的定位结果。实验选取地下目标丰富的场景和实际城市道路场景来测试所提方法的定位性能。实验结果表明，与单一使用探地雷达原始回波数据地图的定位方法相比，深度学习辅助的探地雷达定位方法能避免直接计算原始雷达回波数据间的相似度，减少数据计算量以及数据传输量，具有实时定位能力，同时特征指纹地图对回波数据的变化有鲁棒性，因此所提方法的平均定位误差减少约70%。深度学习辅助的探地雷达定位方法可作为未来自动驾驶车辆在恶劣环境下感知定位方法的补充。

A Deep Learning Assisted Ground Penetrating Radar Localization Method

Under harsh conditions, such as rain, snow, dust, strong light, and dark night, the vision and laser sensors commonly used in autonomous driving solutions may fail because they can not accurately sense the external environment. Therefore, a method for vehicle localization using underground target features sensed by deep learning assisted ground penetrating radar is proposed in this paper. The proposed method is divided into two phases: offline mapping phase and online localization phase. In the offline mapping phase, the ground penetrating radar is used to collect the echo data from the underground targets first, then the Deep Convolutional Neural Network (DCNN) is utilized to extract the target features from the collected echo data, and the extracted target features are saved with the current geographic location information to form a fingerprint map of underground target features. In the localization phase, the DCNN is used to extract the target features from the current echo data collected by the ground penetrating radar first, and then the target feature most similar to the current extracted target feature in the fingerprint map of underground target features is retrieved based on the particle swarm optimization method, and the geographic location information of the retrieved feature is marked as the vehicle localization result by the ground penetrating radar. Finally, the Kalman filter is used to fuse the ground penetrating radar localization result and the mileage information measured by the ranging wheel to obtain a high-precision localization result. The localization performance of the proposed localization method is tested on the experimental scenario with rich underground targets and the actual urban road scenario. The experimental results show that, compared with the single raw data-based ground penetrating radar localization method, the deep learning assisted ground penetrating radar localization method can avoid directly calculating the similarity between the raw radar data, reduce the amount of data computation and data transmission, and has the real-time localization capability. At the same time, the fingerprint map of underground target features is robust to the changes of the raw radar data, so the average localization error of the proposed method is reduced by about 70%. The deep learning assisted ground penetrating radar localization method can be used as a supplement to the detection and localization method of autonomous vehicles in harsh environments in the future.