MonoTemp3D: Lightweight Short-Horizon Temporal Enhancement for Streaming Monocular 3D Car Detection

Monocular 3D object detection (M3OD) is attractive for autonomous driving and resource-constrained robotic platforms due to its low hardware cost. While temporal cues can alleviate depth ambiguity, dense feature matching and explicit motion estimation introduce substantial latency and limit low-latency deployment. To alleviate this burden, we introduce MonoTemp3D, a lightweight dual-level temporal enhancement framework tailored to causal streaming query-based monocular 3D car detection. It efficiently exploits short-term temporal continuity and compact historical representations while preserving low-latency inference. To aggregate informative historical features without heavy pixel-wise matching, an asymmetric cross-scale temporal fusion module is proposed to employ sparse deformable attention as an implicit, task-driven temporal sampler. In addition, we introduce a context-aware query updating mechanism to propagate high-confidence historical object queries as instance-level priors and update them using current observations to stabilize 3D detection. Experiments on KITTI Tracking, KITTI 3D and nuScenes show that MonoTemp3D outperforms representative image-based and video-based M3OD methods. MonoTemp3D achieves inference latencies of 34 ms on an RTX 3090 GPU and 43.5 ms on a Jetson Orin NX (TensorRT, FP16).