Dynamic Self-Maintenance Obstacle Costmap Based on Instance Segmentation and Memory Storage Mechanism

Jiawei Zhang, Wenbo Shi, Chengju Liu, Qijun Chen

Abstract: As the key component of graph search path planning algorithms, environmental costmap is widely used in many robot navigation systems. However, during the lifelong operation of robot, various types of obstacles change the environment constantly, which affect the accuracy of costmap and the effectiveness of path planning. To eliminate the negative effects of obstacle changes in lifelong navigation, we propose a novel dynamic self-maintenance obstacle costmap based on instance segmentation and memory storage mechanism. First, we present an instance-level obstacle segmentation model based on laser-vision fusion. The pose, category, and observation information of obstacles are recorded through clustering and recognition. Then, we suggest an update and maintenance approach for obstacle costmap based on memory mechanism. Imitating the Ebbinghaus Forgetting Curve, the costs are updated by different activity levels in real-time, and obstacles are removed or retained in costmap accurately. In comparison to the widely used layered costmaps, experiments show that the proposed approach is able to remove the obsolete obstacles out of sight in time, enhance the accuracy of costmap and improve the effectiveness of lifelong navigation.