MoNet: Motion-Based Point Cloud Prediction Network

Predicting the future can significantly improve safety of intelligent vehicles, which is a key component in autonomous driving. 3D point clouds accurately model information surrounding environment and are crucial for vehicles to perceive scene. Therefore, prediction has great significance be utilized numerous further applications. However, due unordered unstructured, cloud challenging not been deeply explored current literature. In this paper, we propose novel motion-based neural network named MoNet. The idea proposed MoNet integrate motion features between two consecutive into pipeline. introduction enables more capture variations across frames thus make better predictions motion. addition, content introduced spatial individual clouds. A recurrent MotionRNN temporal correlations both features. Moreover, an attention-based align module address problem missing inference Extensive experiments on large-scale outdoor LiDAR datasets demonstrate performance perform applications using predicted results indicate application potential method.