Engineering · 6 min read · April 22, 2026
Vibration Gestures on Furniture via Efficient FPGA Neural Networks
Researchers compress neural networks for gesture recognition on low-power FPGAs, eliminating complex preprocessing and cutting energy use to under 1.2 mJ per inference.
Source: arxiv/cs.AI · Koki Shibata, Tianheng Ling, Chao Qian, Tomokazu Matsui, Hirohiko Suwa, Keiichi Yasumoto, Gregor Schiele · open original ↗ ↗
Compact 1D-CNN models on FPGAs enable real-time vibration-based gesture recognition on furniture with minimal energy and preprocessing overhead.
- — Raw waveform input replaces spectral preprocessing, shrinking input size 21x without accuracy loss.
- — Two lightweight architectures (1D-CNN, 1D-SepCNN) reduce parameters from 369M to 216 while maintaining performance.
- — Integer-only quantization and automated RTL generation enable direct FPGA deployment without manual optimization.
- — Ping-pong buffering in 1D-SepCNN handles tight memory constraints on low-cost Spartan-7 FPGAs.
- — Hardware-aware search framework balances accuracy, latency, energy, and deployability constraints automatically.
- — 6-bit 1D-CNN achieves 97% accuracy with 9.22 ms latency; 8-bit variant reaches 6.83 ms (53x CPU speedup).
- — Both models consume under 1.2 mJ per inference, enabling months of continuous operation on battery.
Frequently asked
- Spectral preprocessing (FFT, filtering) requires complex on-board logic and large intermediate buffers. Raw waveform input shrinks the data 21x and eliminates preprocessing hardware, freeing FPGA resources for the neural network itself. This trade-off works because modern compact CNNs learn features directly from raw signals without sacrificing accuracy.