Fpga Implementation of 1D Convolutional Networks Enables 0.970 Accuracy Vibration-Based Gesture Recognition with 21x Reduced Input

The pursuit of intuitive smart home interfaces drives innovation in non-invasive sensing technologies, and researchers are now focusing on recognising gestures through everyday vibrations. Koki Shibata, Tianheng Ling, and colleagues at the Nara Institute, along with Chao Qian, Tomokazu Matsui, Hirohiko Suwa, and Keiichi Yasumoto, present a significant advance in this field by demonstrating highly accurate gesture recognition directly from furniture surfaces. Their work overcomes limitations of previous systems, which demanded substantial computing power and energy, by deploying compact neural networks on low-power Field-Programmable Gate Arrays (FPGAs). This energy-efficient approach, achieved through streamlined data processing and optimised network architectures, delivers real-time performance with minimal energy consumption, paving the way for truly pervasive and sustainable smart home experiences.

Vibration Sensing Transforms Tables to Interfaces

Scientists have developed Smatable, a system that transforms ordinary tables into touch-sensitive interfaces using vibration sensors. The core idea leverages readily available and inexpensive sensors to detect user interactions, opening possibilities for interactive displays, gaming, and control systems. The system employs an array of vibration sensors beneath the table surface. Sophisticated algorithms process signals from these sensors, accurately identifying the location and type of touch interaction.

Implementing this signal processing and control logic on FPGAs achieves low latency, energy efficiency, and real-time performance. Researchers utilise techniques like quantization and network pruning to reduce the computational complexity of Deep Learning models, making them suitable for resource-constrained embedded systems. This work demonstrates the feasibility of using vibration sensors for accurate touch detection on large surfaces. The team employs 1D Convolutional Neural Networks (CNNs) for feature extraction and classification of touch events, utilising Depthwise Separable Convolutions to reduce model size and processing time.

Integer Quantization converts calculations to integer arithmetic, enabling efficient implementation on FPGAs. The entire system is successfully implemented on an FPGA, achieving real-time performance and low power consumption, facilitated by the ElasticAI Framework for design space exploration and optimisation. Experiments demonstrate high accuracy in detecting and classifying touch events. The FPGA implementation achieves significant performance improvements compared to software-based approaches, exhibiting low latency and power consumption, making it suitable for interactive applications. The research highlights the importance of energy efficiency for pervasive computing and edge computing applications.

Raw Waveform Input for Efficient Gesture Recognition

Scientists have developed a novel approach to vibration-based gesture recognition, prioritising energy efficiency and real-time performance for smart home applications. The study addresses limitations in previous work that relied on complex signal processing and computationally expensive neural networks. Researchers engineered a system that bypasses traditional spectral preprocessing, directly inputting raw waveform data, achieving a 21-fold reduction in input size without compromising accuracy. These networks dramatically reduce the number of parameters, while maintaining comparable accuracy to more complex models. Scientists implemented integer-only quantization and automated Register Transfer Level (RTL) generation to facilitate seamless FPGA deployment. A ping-pong buffering mechanism was integrated into the 1D-SepCNN architecture to improve deployability under tight memory constraints.

The study pioneered a hardware-aware search framework that enables constraint-driven model configuration selection, considering accuracy, deployability, latency, and energy consumption. Experiments employed an AMD Spartan-7 XC7S25 FPGA and utilised swipe-direction datasets with multiple users and ordinary tables. Results demonstrate that a 6-bit 1D-CNN achieves 0. 970 average accuracy with 9. 22ms latency, while an 8-bit 1D-SepCNN further reduces latency to 6.

83ms, representing over a 53-fold speedup compared to CPU-based processing, with slightly lower accuracy of 0. 949. Both configurations consume under 1. 2 mJ per inference, demonstrating suitability for long-term edge operation.

Low-Power Gesture Recognition via Waveform Analysis

Scientists have achieved a breakthrough in energy-efficient gesture recognition using vibration sensing, demonstrating real-time performance on low-power hardware. The team replaced complex signal processing with direct analysis of raw waveform data, reducing input size by a factor of 21 without compromising accuracy. This simplification significantly lowers computational demands and energy consumption. To further enhance efficiency, researchers designed two lightweight neural network architectures, 1D-CNN and 1D-SepCNN, tailored for embedded FPGAs.

These networks dramatically reduced the number of parameters, while maintaining comparable accuracy to more complex models. Experiments revealed that a 6-bit 1D-CNN achieved 0. 970 average accuracy across multiple users, with a low latency of 9. 22 milliseconds. An 8-bit 1D-SepCNN further improved performance, reducing latency to 6.

83 milliseconds, over 53times faster than CPU-based inference, with slightly lower accuracy of 0. 949. Crucially, both networks consumed less than 1. 2 millijoules per inference, demonstrating suitability for long-term operation in energy-constrained environments. The team achieved this efficiency through integer-only quantization and automated hardware generation, enabling seamless deployment on the Spartan-7 XC7S25 FPGA. A novel ping-pong buffering mechanism in the 1D-SepCNN architecture further improved performance under tight memory constraints. This work delivers a pathway to practical, low-power smart home interfaces.

Low-Latency Gesture Recognition on Embedded Systems

This research demonstrates a fully functional, FPGA-deployable system for vibration-based gesture recognition, designed for low-latency and energy-efficient interaction with smart surfaces. The team achieved this by streamlining the process, eliminating complex signal processing steps and instead operating directly on raw vibration data, reducing input size significantly without compromising accuracy. Two lightweight convolutional neural network architectures, 1D-CNN and 1D-SepCNN, were developed, dramatically reducing the number of parameters needed for effective gesture recognition and enabling deployment on resource-constrained hardware.