LMTSDet: a lightweight Mamba-transformer detector with spatial-spectral fusion and feature calibration for real-time traffic sign detection

Traffic sign detectors must preserve small-object detail and scene context under strict latency and memory constraints. We developed LMTSDet, an RT-DETR-based detector that uses the published MobileMamba backbone and three task-specific components. The Spatial-Spectral Fusion Module combines state-space spatial modeling with frequency-domain modulation, the Feature Calibration Module recalibrates directional spatial responses, and the Spatial Calibration Fusion Network aggregates calibrated multi-scale features. Structured channel pruning, knowledge distillation, and fine-tuning further reduce the deployment cost. On TT100K, LMTSDet achieves 92.8% [email protected] with 12.1M parameters and 28.6 GFLOPs. The compressed model retains 91.5% [email protected] and reaches 156 FPS on an NVIDIA Jetson AGX Orin with TensorRT FP16. On GTSDB, direct transfer and target-domain fine-tuning yield 87.3% and 93.0% [email protected], respectively. These results indicate that LMTSDet provides a favorable accuracy-efficiency trade-off across the evaluated settings, although performance on adverse-weather and more diverse regional datasets remains to be established.