This study addresses the decline in detection precision caused by sparse point clouds at medium and long ranges in 3D object detection. An improved range-aware CenterFormer framework, termed DA-CenterFormer, is proposed. Specifically, the original backbone network is replaced with a UNetV2 architecture incorporating deformable convolutions, and a range-aware IoU-weighted loss function is designed to improve long-range detection performance. Experimental results on the Waymo open dataset show that the proposed method significantly improves vehicle and pedestrian detection accuracy at ranges of 30–50 m and beyond 50 m.