Dynamic Occupancy Grid Map Update Method using Camera and LiDAR Object Detection for Autonomous Driving

For safe autonomous driving, cameras are widely utilized to detect and classify objects, whereas light detection and ranging (LiDAR) provides accurate distance information for objects. In this study, we employed cameras and LiDAR to obtain three-dimensional (3D) bounding box information, such as class information, and the position and orientation of the objects using deep learning-based recognition techniques. Among various methods for representing the surrounding environment, the dynamic occupancy grid map (DOGM) has the advantage of representing the dynamic and static states of each object. However, utilizing only the DOGM increases the difficulty of considering the class information of each object. Herein, we utilized 3D bounding box information as input values to construct DOGM, including semantic information. Specifically, we applied the Dempster-Shafer theory of evidence-based combination rules to update the semantic information of the grid cell. Additionally, the developed methods are verified using the nuScenes public data sets, including camera and lidar measurements.