Accelerator-Free Thermal Pose Estimation: Accuracy–Efficiency Trade-Offs on the Raspberry Pi 5

Deploying human pose estimation on low-cost, accelerator-free hardware remains a key open question for smart home activity recognition, where affordability and wide availability often matter as much as raw inference speed. This work addresses that question by evaluating thermal-image pose estimation on a Raspberry Pi 5, a CPU-only single-board computer with no dedicated neural-network accelerator. Building on a previously collected single-person thermal dataset of 2500 images annotated with 17 body keypoints, YOLOv8-pose, YOLO11-pose, and YOLO26-pose models were trained across five model scales (n–x) at three input resolutions (640×512, 320×256, and 160×128 px) and deployed on the Raspberry Pi 5. Models were evaluated in native PyTorch, ONNX, and NCNN export formats, with per-image latency, power consumption, energy per frame, and pose mAP50–95 measured for each configuration. Pose accuracy scaled consistently with model size and input resolution, reaching up to 95.6% pose mAP50–95 for YOLO26-x at 640×512 px, with diminishing returns for the largest model scales. Export-format conversion to ONNX or NCNN preserved pose accuracy almost exactly relative to native PyTorch, with mean degradation below 0.6 percentage points, in contrast to the accuracy loss associated with numerical-precision reduction on GPU-accelerated platforms. NCNN consistently achieved the lowest per-image latency and energy consumption among the three formats, delivering up to 2.1–4.3× speed-up over PyTorch while keeping power draw up to 10 W across all configurations. The findings indicate that, unlike GPU-based edge deployment where precision reduction is the dominant efficiency driver, export-format optimization governs the latency–power–energy trade-off on CPU-only hardware, and that NCNN-based deployment on low-cost, widely available single-board computers offers a practical alternative for smart home thermal pose estimation.