ExistenceMap-PointPillars: A Multi-Fusion Network for Stable 3D Object Detection with Pseudo 2D Maps

Recognition of surrounding traffic participants is important for safe driving of automated vehicles. There are approaches to recognize the surrounding environment that are 3D object detection using only a LiDAR sensor or fusion with a camera. However, when information from the other sensor is lost, stable object detection becomes difficult in the fusion method. In this paper, we propose ExistenceMap-PointPillars, which is an effective method for 3D object detection with information from other sensors by a simple modification to LiDAR base 3D object detection network. Pseudo 2D map which shows object existence regions can be generated from any information and can be easily fused with the 3D object detection network. The experimental results on our dataset show that ExistenceMap-PointPillars improves mean Average Precision (mAP) by +3.59% over PointPillars and outperforms other sensor fusion methods. ExistenceMap-PointPillars is effective in significantly reducing False Positives and enables stable 3D object detection even when information from other sensors is intercepted.