Biomechanical Posture Analysis System Using Computer Vision: An Edge-Computing Architecture Integrating Finite State Machines and Large Language Models

Abstract

Traditionally, computer vision integration in fitness applications has relied on cloud-based processing or simple motion detection, which frequently jeopardizes user privacy and does not uphold stringent biomechanical standards. A novel edge-computing architecture for real-time posture correction and repetition tracking is presented in this paper. The system extracts three-dimensional topological information from standard Red-Green-Blue (RGB) video feeds using a lightweight 33-landmark pose estimation model (MediaPipe BlazePose). We put in place a deterministic Finite State Machine (FSM) powered by dynamic Euclidean geometric angle computations to guarantee exercise effectiveness and avoid injury. This layer filters out momentum-based lifting behaviours and incomplete repetitions while rigorously enforcing a full range of motion (ROM). Additionally, we incorporate a local Meta Llama 3 Large Language Model (LLM) instance that uses real-time performance metrics to provide customized, JavaScript Object Notation (JSON)-structured workout feedback. Our "Offline Edge AI" method, according to experimental results, maintains a processing latency of less than 45 ms and achieves a repetition counting accuracy of 85%, demonstrating that advanced biomechanical analysis is possible without the high bandwidth and privacy risks associated with cloud-based alternatives.

Keywords

Graphical Abstract

Novelty Statement

An edge-based framework integrating MediaPipe BlazePose, FSM, and local LLM enables privacy-preserving, real-time biomechanical posture analysis and personalized coaching.