Energy-per-Pixel Analysis on Edge-TPU-Based Aerial Object Detection on MCU-Class Devices

Deploying object detection models on low-power embedded devices requires a careful balance between detection accuracy, runtime performance, and energy efficiency, especially for aerial images that contain small objects. We present a hardware-measured evaluation of Edge-TPU-compatible detectors on an MCU-class NXP i.MX RT1176 (bare-metal FreeRTOS, Cortex-M7) driving a Coral Edge TPU over an internal USB 2.0 link. We benchmark quantized YOLOv5-based detectors at input resolutions from 256 to 1024 px, reporting the accuracy (size-stratified AP_S on the VisDrone person split) together with the measured per-frame and per-pixel energy, sampled at 100 Hz with an automatic window and N ≥ 5 repetitions. We use the energy per pixel as a resolution-normalized view, not as a new metric, to show where the analytical model fails. The measured per-pixel energy departs from the compute (MAC)-based model because the inference is transfer-bound: it is dominated by the instruction stream and input activations re-sent over USB on every invoke, while the parameters stay cached on-chip. An Edge TPU clock sweep and a compiler-byte analysis agree independently that about 79% of each invocation is the USB transfer. This gives concrete design rules (a lower clock is more energy-efficient; ReLU is preferred to SiLU), and shows that the USB 2.0 ceiling is architecturally fundamental for the MCU class.