Carbon-Aware Edge Shrinkage: Dynamic Head Pruning and Adaptive Voltage-Frequency Scaling for Green Object Detection

The ubiquitous deployment of high-capacity deep vision detectors across millions of autonomous edge devices poses severe energy and environmental challenges. While contemporary model compression techniques focus almost exclusively on static parameter pruning or post-training quantization, they remain insensitive to real-time environmental context, battery state of charge (SoC), and transient grid carbon intensity. In this paper, we propose a closed-loop Green AI framework, termed Carbon-Aware Edge Shrinkage (CAES), that dynamically modulates YOLO neural network capacity and hardware operating states. Our architecture orchestrates real-time structural pruning of redundant multi-scale detection heads and convolutional channels guided by instantaneous NVML power measurements and thermal envelopes. Concurrently, an adaptive governor modulates GPU Dynamic Voltage and Frequency Scaling (DVFS). Empirical evaluations on NVIDIA Jetson AGX Orin platforms demonstrate that CAES slashes per-frame energy consumption from 4.85 J down to 2.35 J (a 51.5% reduction) and lowers operational carbon emissions by 51.6%, while preserving over 95.8% of nominal detection accuracy (0.602 vs 0.628 mAP@50:95). Furthermore, CAES mitigates thermal throttling by 17.2°C, establishing an optimal Pareto frontier between sustainability and safety-critical edge perception.