PoseQueue : leveraging in-memory time-series databases for real-time pose estimation in healthcare applications
Kauppinen, Vy (2025)
Katso/
Sisältö avataan julkiseksi: 11.05.2027
Diplomityö
2025
School of Engineering Science, Tietotekniikka
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe20251111107109
https://urn.fi/URN:NBN:fi-fe20251111107109
Tiivistelmä
Artificial intelligence (AI) is increasingly applied in digital healthcare, particularly in physiotherapy and rehabilitation, where 3D pose estimation enables real-time assessment of human movement. However, to overcome hardware limitations, current systems often rely on a conventional client–server design, where the mobile application sends each video frame to the AI server for inference and awaits the result before continuing. This sequential exchange introduces high latency, causes dropped frame rate when network conditions fluctuate, and limits scalability as user concurrency increases.
This thesis presents PoseQueue, a Redis-based in-memory buffering architecture developed to address these limitations and enhance AI-assisted medical applications. Unlike the baseline client-server design, PoseQueue decouples client requests from the inference process by caching pose estimation results as time-series data in memory. This enables the server to maintain continuous processing even if client frames arrive irregularly, minimising redundant computation, stabilising throughput under load, and significantly reducing response latency.
Performance metrics, including throughput, latency, failure rate, accuracy, CPU and memory utilisation, and energy efficiency, were evaluated under controlled stress testing. Results show that PoseQueue achieved a 39% increase in efficiency rate compared to the baseline system, with lower failure rates, faster response times, and higher accuracy in exercise repetition counting. The findings demonstrate that in-memory architectures can substantially improve the scalability, reliability, and usability of AI servers for real-time medical applications. Beyond immediate benefits for the partner company’s physiotherapy service, Pose-Queue also offers a foundation for future research, including the use of buffered pose data to retrain estimation models for greater adaptability in clinical settings.
This thesis presents PoseQueue, a Redis-based in-memory buffering architecture developed to address these limitations and enhance AI-assisted medical applications. Unlike the baseline client-server design, PoseQueue decouples client requests from the inference process by caching pose estimation results as time-series data in memory. This enables the server to maintain continuous processing even if client frames arrive irregularly, minimising redundant computation, stabilising throughput under load, and significantly reducing response latency.
Performance metrics, including throughput, latency, failure rate, accuracy, CPU and memory utilisation, and energy efficiency, were evaluated under controlled stress testing. Results show that PoseQueue achieved a 39% increase in efficiency rate compared to the baseline system, with lower failure rates, faster response times, and higher accuracy in exercise repetition counting. The findings demonstrate that in-memory architectures can substantially improve the scalability, reliability, and usability of AI servers for real-time medical applications. Beyond immediate benefits for the partner company’s physiotherapy service, Pose-Queue also offers a foundation for future research, including the use of buffered pose data to retrain estimation models for greater adaptability in clinical settings.