Time sequence fusion point cloud 3D target detection method and system, terminal and medium

The invention provides a time sequence fusion point cloud 3D target detection method and system, a terminal and a medium. The time sequence fusion point cloud 3D target detection method comprises the steps of obtaining point cloud data of a time sequence; aligning the point cloud data to the same coordinate system; in the training process, data enhancement with consistent time sequence is used for training to solve the problem of non-uniform object distribution; after encoding the point cloud into an aerial view feature map, dynamically fusing features of past moments for the feature map of the current frame by using a deformable attention mechanism; and sending the fused feature map into a detection head to predict an object. According to the method, the detection performance is remarkably enhanced, and the method can be suitable for any aerial view detection method and can be expanded to a time sequence with any length. 本发明提供一种时序融合的点云3D目标检测方法、系统、终端及介质，包括：获取一时序的点云数据；将所述点云数据对齐到同一个坐标系下；在训练过程中，使用时序一致的数据增强进行训练来解