Abstract: With the advancement of autonomous driving technology, 3D object detection based on LiDAR point clouds has become a critical task. However, existing methods mostly rely on single-frame detection, failing to fully utilize temporal information, which leads to lower detection accuracy, occlusion issues, and object loss. a novel multi-frame object detection model that integrates temporal information to improve detection performance is proposed. To effectively leverage multi-frame data, the proposed model introduces a matching inference module and a scene flow module. The matching inference module predicts the object's direction and velocity based on its initial position and updates its subsequent positions. The scene flow module integrates each point’s position, velocity, and direction in the point cloud, providing more accurate data for inference. Experiments show that the proposed model achieves 59.2% mAP and 67.4% NDS on the nuScenes dataset, as well as 89.03% average precision for the car category in the moderate difficulty evaluation on the KITTI dataset, outperforming most object detection models. Additionally, we visualized the detection results and conducted extensive ablation studies, further validating the model's effectiveness and the feasibility of the proposed improvements.