Approaching the Real-World: Supporting Activity Recognition Training with Virtual IMU Data
summary
System converts YouTube videos into 3D motion tracking and generates virtual IMU streams, then applies adaptive video selection to remove non-human objects, occlusions, motion blur and extreme camera motion. Evaluated on free‑weight exercise classification (13 classes) using a 3.5‑minute real IMU dataset versus up to 100× larger virtual dataset (≈3.2 hours per activity), training deep models that improve overall F1 from about 0.5→0.75, showing the pipeline produces usable virtual sensors and boosts performance.
AI-written summary · about
related talks
06mIMUTube: Automatic extraction of virtual on-body accelerometry from video for human activity ...
13mIMUCoCo: Enabling Flexible On-Body IMU Placement for Human Pose Estimation and Activity Recognition
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
22mRunning, walking, sitting or biking? - Motion prediction with acceleration and rotation
27mDissecting ActivityRecognition with Multi-Sensor Simulation and Its Practical Deployment
06mDeepMV: Multi-View Deep Learning for Device-Free Human Activity Recognition
27mAnalysis of sports scenes using computer vision models
28mEnhancing Upper Limb Rehabilitation with Unsupervised Machine Learning Models