Abstract: Aiming at the limitation of single sensor of intelligent vehicle in perceiving environment, an environment perception algorithm based on multi-source heterogeneous information fusion with GPS, binocular depth camera and 16-line LIDAR was proposed, so as to solve the limitation of single sensor in environment perception. Based on GPS time synchronization, the proposed algorithm took the calibrated data of a single sensor as the input of an improved joint calibration algorithm, so as to solve the optimal spatial transformation matrix of multi-source sensors. The SVM segmentation point cloud and European clustering were used to extract the geometric features of the point cloud information, so as to obtain the location information of the obstacles. The space transformation matrix was used to transform the obstacle point cloud information into the image coordinate system, and the 3D point cloud was mapped to the 2D point cloud information. The obstacle position information and category information in images solved by deep learning algorithm were fused to achieve target detection. Through the verification of KITTI dataset and real vehicle test, the accuracy of the proposed algorithm is between 78.13%~85.56% and the average detection time of each frame is 0.16~0.19 s. The proposed algorithm can effectively detect the obstacle both in the light changing and occlusion environment, which has a promising engineering application prospect.

Key words: vehicle engineering; intelligent vehicle; time synchronization; joint calibration; information fusion; environment perception

摘要： 针对智能汽车单一传感器环境感知的局限性，提出一种基于GPS、双目深度相机、16线激光雷达等多源信息融合的环境感知算法，解决单一传感器环境感知的局限性问题。算法在GPS时间同步基础上，将单一传感器标定后的数据作为改进联合标定算法输入，求解多源传感器最优空间变换矩阵。使用SVM分割点云及欧式聚类提取点云信息的几何特征，获取障碍物的位置信息；利用空间变换矩阵将障碍物点云信息变换到图像坐标系，并将3D点云映射为2D点云信息；融合基于深度学习算法求解的图像中障碍物位置信息与类别信息，实现目标检测。经KITTI数据集及实车测试验证，该算法准确率在78.13%~85.56%之间，每帧数据检测平均耗时0.16~0.19 s，在光照变化、目标遮挡环境下均能有效的进行目标检测，具有较好的工程应用前景。

关键词: 车辆工程；智能汽车；时间同步；联合标定；信息融合；环境感知