Teaching RF to Sense without RF Training Measurements
summary
Paper presents a pipeline that generates RF training data from online videos and uses it to train a WiFi-based gym activity classifier without any real RF measurements. It translates YouTube videos into simulated WiFi CSI by extracting and aligning 3D meshes, applies Born approximation to simulate reflections from body points, computes spectrogram features, and trains a neural net. Evaluation on 10 subjects, 3 rooms, 1543 repetitions yields 86% accuracy on multi-repetition windows and 81% on single repetitions, matching state of the art.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:39ML for RF Sensing
- 1:40Training RF Sensing Systems without Training Data
- 2:13Case Study: Gym Activity Classification
- 2:313D Mesh Extraction
- 2:593D Mesh Alignment (cont.)
- 3:27Sample Aligned 3D Meshes
- 4:19Video-to-RF Simulation Validation
- 4:32Training Gym Activity Classifier
- 4:57Performance Evaluation
- 5:22Conclusions
related talks
17mSLNet: A Spectrogram Learning Neural Network for Deep Wireless Sensing
05mThe StoryTeller: Scalable Building- and AP-independent Deep Learning-based Floor Prediction
12mAdapting Wireless Mesh Network Configuration from Simulation to Reality via Deep Learning
24mAn Open Channel Identifier using GNU Radio
01h 13mHealth Monitoring with Wireless Sensors and Machine Learning
42mTeaching the iPhone to See
06mIMUTube: Automatic extraction of virtual on-body accelerometry from video for human activity ...
32mMan or Machine?: Developing a "Turing Test" for Radio Intelligence