REAL-TIME POSE EVALUATION USING AN OPTIMIZED BLAZEPOSE-LITE MODEL FOR LOW-RESOURCE DEVICES

0 1

Abstract

This study proposes an optimized real-time pose evaluation system based on the BlazePose-Lite model, specifically adapted for low-resource devices such as smartphones, Raspberry Pi boards, and low-end laptops. The relevance of this research is driven by the growing need for real-time human pose estimation in fitness applications, rehabilitation systems, mobile health monitoring, and embedded AI solutions, where computational resources are often limited. The primary aim of the study is to enhance the inference speed of BlazePose-Lite while preserving pose-estimation accuracy. The methodology includes a multi-stage optimization pipeline: TensorFlow Lite conversion using FP16 and INT8 quantization, structured model pruning, graph simplification through operator fusion, and temporal smoothing via Exponential Moving Average and Kalman filtering. The optimized model was evaluated on several low-resource platforms, where performance was measured using FPS, latency, CPU load, RAM usage, and keypoint accuracy metrics (PCK and RMSE). Experimental results show that INT8-quantized BlazePose-Lite achieves a 2×–3× increase in inference speed, reaching up to 26–32 FPS on mid-range smartphones and 12–16 FPS on Raspberry Pi 4, while model size was reduced by up to 75%. Accuracy loss remains within 1–3%, making the optimized model suitable for real-time applications. The practical significance of the study lies in enabling robust, efficient, and deployable human pose-tracking systems for IoT fitness devices, mobile coaching applications, and embedded smart health platforms.