In the field of intelligent transportationthe real-time three-dimensional(3D) object detection in road scenes is of great significance to ensure the safety of vehicle driving.The use of radar point cloud and image data fusion is able to achieve the effect of complementary advantages.Howeverin order to achieve higher detection accuracythe 3D object detection algorithms using radar and image data fusion usually adopt the two-stage networkbut the computing speed is slower than that of the one-stage networkand the speed of detection system is very important in practical application.Aiming at the above problemsa real-time detection method of 3D object is designed based on the improvement of one-stage network RetinaNet.The 3D anchors are mapped to the feature maps of the point cloud and image respectivelyand ROI pooling is used to convert the regions cropped by the anchor on the feature map into the same size and fuse them.Finallythe regression parameters and categories of the target boundary box are outputand the target prediction boundary box is obtained by adjusting the anchor box.Experiments on KITTI dataset show that the proposed network is superior to the comparison algorithms in the accuracy and time consumption of pedestrian and vehicle detection.