SimpleYOLO Detects Driver Behavior on Edge Hardware
A lightweight object detector reports 95.46% mAP@0.5 on a driver monitoring benchmark while running on a low-power edge platform.
Updated on Nov. 14, 2026 in Artificial Intelligence

Researchers have proposed SimpleYOLO, an ultra-lightweight object detector designed for edge-based driver monitoring. The model reports 95.46% mAP@0.5 on 6,284 images from the driving-fvsg3 DMS benchmark.
Why it matters
Existing lightweight detectors often rely on aggressive network scaling, which weakens feature representation and reduces robustness for fine-grained driver behaviors. The work targets a tradeoff between accuracy and the compute budget available to in-vehicle hardware.
SimpleYOLO has 0.35M parameters and 2.1 GFLOPs, and its INT8 kmodel is 0.844 MiB with a mean KPU latency of 64.96 ms on the Canaan K230 platform. Whole-board power measured 1.43 W, though the stub does not compare these figures against competing detectors on identical hardware.
The details
The backbone is rebuilt with Ghost-based feature generation to cut computational redundancy. A Multi-dimensional Compensatory Attention Module (MCAM) applies spatial, height-wise, width-wise and channel gates to partially recover accuracy lost to compression.
Roadmap
The work builds on the YOLO family of real-time detectors, reducing parameter count and compute to fit low-power edge chips. It reflects a broader push to move driver monitoring from cloud or high-end hardware into the vehicle itself.
Driver monitoring systems built on compact detectors could run on inexpensive in-vehicle hardware rather than requiring large computing modules. The reported power draw of 1.43 W points to feasibility for devices that operate continuously inside a car.
The takeaway
Lightweight architectures that trade some raw capacity for efficiency may be the practical path for safety features on constrained hardware. Readers tracking in-car AI should watch whether these benchmark gains hold under real driving conditions.
Further reading
For related coverage, explore Artificial Intelligence on the site.
Source note: This article includes information reported by Nature.







