Abstract: ‍ ‍With the development of artificial intelligence technology and the in-depth study of target-tracking theory， target-tracking technology has become widely used in real life. Target tracking is an essential technology for many vision applications. Although the target tracking research for two-dimensional images has already achieved fruitful results， target-tracking technology based on a three-dimensional point cloud is still in the research and development stage. Target-tracking technology based on a lidar point cloud has advanced somewhat in recent years thanks to the widespread use of lidar and development of deep-learning technology in the area of 3D point clouds. The existing tracking algorithms based on lidar point cloud targets are mainly divided into two categories： traditional filtering algorithms and deep-learning algorithms. Although point-cloud target-tracking algorithms based on traditional filtering can achieve better results， it is difficult to assign optimal parameters to such algorithms， and these algorithms can easily fail when dramatic changes occur in a scene. Most of the lidar point-cloud target-tracking algorithms based on deep learning utilize "detection-track" architectures. The biggest problem with this architecture is that the back-end tracking task depends heavily on the front-end detection results， and the back-end tracker cannot track when the front-end detector fails. This creates many target-loss issues. This paper introduces a deep-learning architecture centered on motion prediction to solve these problems. The architecture combines detection and motion prediction， which is mainly divided into two stages. The object is located in successive frames in the first step after being detected using point cloud feature extraction， segmented from the point cloud， and located. In the second stage， the motion prediction update branch is used to optimize the target box to obtain a more accurate target position. The results of experiments showed that this approach is successful and better able to handle dramatic shifts in the scene than the conventional filtering approach. Compared with the "detection-track" architecture used in deep-learning methods， it reduces the target losses. Better accurate and robust tracking results could be obtained in lidar point-cloud target tracking.