IMUCoCo: Enabling Flexible On-Body IMU Placement for Human Pose Estimation and Activity Recognition
summary
Framework maps IMU time series from any skin location into a unified feature space using XYZ spatial coordinates plus a region category and feeds that into downstream pose and activity models. Training uses synthetic IMU signals extracted from videos at thousands of body mesh points. Evaluations use dense IMU captures then test 3–6 real sensors and show global joint-angle error stays comparable across typical and atypical placements, and the system also recommends sensor layouts.
AI-written summary · about
chapters (8)
- 0:00Introduction and setup
- 1:59IMU sensing challenges
- 4:46Synthetic IMU data generation
- 7:15Adapting to sensor placement
- 8:38Model performance results
- 9:42Placement suggestions for activities
- 10:44Future applications and ADHD study
- 11:27Q&A and conclusion
related talks
06mApproaching the Real-World: Supporting Activity Recognition Training with Virtual IMU Data
06mIMUTube: Automatic extraction of virtual on-body accelerometry from video for human activity ...
24mProject GaitNet: Ushering in the ImageNet moment for human Gait kinematics
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
10mProactive Sensing for Improving Hand Pose Estimation
13mIMU Sensor ML Action Detection Optimization with Custom Instructions on Embedded RISC-V Cores
07mMultimodal Emotion Recognition of Hand-Object Interaction
31mData as an instrument for more efficient collaboration with robots