The new generation of 4D imaging radar not only offers all-weather capabilities and low cost but also provides high-resolution point cloud. Its emergence presents novel solutions for autonomous driving perception systems. To address the disorder and continuity inherent in point clouds, we propose a 3D point cloud object detection method based on temporal graph neural networks. First, graphs are adopted as the representation format for point cloud data, perfectly aligning with the geometric structure of point clouds. Second, graph neural networks are used as the feature extraction network to extract spatial features from each frame of point cloud. These features are then input into a gated recurrent unit alongside hidden features from the previous frame to extract temporal features. The spatiotemporal fused features output by the gated recurrent unit are used both for object detection in the current frame and as hidden features for the next frame. Finally, this method was validated on the public TJ4DRadSet dataset, achieving performance of 43.68% and 38.43% on the mAP3D and mAPBEV, respectively.