Answers are grounded in this paper and cite their sources.
Driver monitoring systems (DMSs) impose strict computational and memory constraints on edge devices, making real-time perception challenging. Existing lightweight detectors often rely on aggressive network scaling, which weakens feature representation and reduces robustness for fine-grained driver behaviors. To address this issue, we propose SimpleYOLO, an ultra-lightweight object detector for edge-based driver monitoring. The backbone is reconstructed using Ghost-based feature generation to reduce computational redundancy, while a Multi-dimensional Compensatory Attention Module (MCAM) is introduced to partially compensate for accuracy loss under excessive compression. MCAM applies multiplicative spatial, height-wise, width-wise, and channel gates. Experiments on 6,284 images from the public driving-fvsg3 DMS benchmark show that SimpleYOLO achieves 95.46% mAP@0.5 with only 0.35M parameters and 2.1 GFLOPs. The INT8-deployed model on the Canaan K230 RISC-V edge platform has a 0.844-MiB kmodel and a mean KPU latency of 64.96 ms. Under the detector-core timing protocol, it achieves 10.32 FPS. Its measured whole-board power is 1.43 W, corresponding to 7.22 FPS/W.
No takes yet. Share an insight, caveat, or question.
Fu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: