9308-GRP On-Device Exercise Recognition and Repetition Count from Real-Time Video Stream Data

Cloud-based fitness trackers send private video off the phone. We present a fully on-device pipeline that recognises eight activities, including complex exercises such as hanging leg raise, squat shoulder press and skipping, and counts repetitions in real time from the live camera or a gallery video. YOLOPose-26 model’s 17-point 2D body pose feeds pose features into a lightweight temporal classifier. We compare three models with matched budgets of ~126k parameters: a Bi-LSTM, a Transformer and an ST-GCN. Under 5-fold cross-validation on 185 clips, the Transformer is the most accurate, reaching 98.1% window accuracy and 97.8% macro-F1. All three run as fp16 Core ML models under 0.5 MB in under 0.3 ms per window.

Authors

Institutions

Publication Details

Journal
DigitalCommons - Kennesaw State University (Kennesaw State University)
Published
2026-11-20
Primary Topic
Human Pose and Action Recognition
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

9308-GRP On-Device Exercise Recognition and Repetition Count from Real-Time Video Stream Data

ARNAB JOY
DigitalCommons - Kennesaw State University (Kennesaw State University)
Human Pose and Action Recognition
article

9308-GRP On-Device Exercise Recognition and Repetition Count from Real-Time Video Stream Data

ARNAB JOY
article en

Abstract

Cloud-based fitness trackers send private video off the phone. We present a fully on-device pipeline that recognises eight activities, including complex exercises such as hanging leg raise, squat shoulder press and skipping, and counts repetitions in real time from the live camera or a gallery video. YOLOPose-26 model’s 17-point 2D body pose feeds pose features into a lightweight temporal classifier. We compare three models with matched budgets of ~126k parameters: a Bi-LSTM, a Transformer and an ST-GCN. Under 5-fold cross-validation on 185 clips, the Transformer is the most accurate, reaching 98.1% window accuracy and 97.8% macro-F1. All three run as fp16 Core ML models under 0.5 MB in under 0.3 ms per window.

DigitalCommons - Kennesaw State University (Kennesaw State University)
Kennesaw State University (US)
Human Pose and Action Recognition
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon