Efficient Thermal Pose Estimation: Balancing Accuracy and Edge Deployment for Smart Home Activity Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Edge Computing

Background:

• Smart home activity recognition requires efficient human pose estimation.
• Edge deployment poses constraints on computational resources and power.

Purpose of the Study:

• To investigate efficient thermal-image human pose estimation for edge devices.
• To evaluate YOLO models' performance across various scales and resolutions.
• To assess the impact of different precisions (FP32, FP16, INT8) on accuracy and runtime.

Main Methods:

• Collected and annotated a 2500-image single-person thermal dataset with 17 keypoints.
• Trained and evaluated YOLO11-pose and YOLOv8-pose models at multiple scales and input resolutions (640x512, 320x256, 160x128).
• Measured accuracy using mAP50-95 and OKS, and runtime performance (latency, power) on NVIDIA Jetson platforms with PyTorch and TensorRT.

Main Results:

• Pose accuracy decreased with reduced input resolution.
• TensorRT FP16 maintained pose accuracy while significantly reducing latency and power consumption compared to FP32.
• INT8 offered further power savings and potential latency improvements but introduced configuration-dependent accuracy losses.

Conclusions:

• FP16 precision with TensorRT provides an optimal accuracy-efficiency trade-off for thermal pose estimation on edge devices.
• Practical implementation depends on specific edge device capabilities and memory constraints.
• This research informs efficient AI model deployment for real-time applications in constrained environments.