A Lightweight and Low-Latency Framework for Tactile Texture Recognition at Edge Using Linear Predictive Coding (LPC)

Tactile sensing is a critical component of the Internet of Things (IoT) and robotic systems, enabling machines to perceive physical properties such as surface texture. However, deploying real-time texture recognition on edge devices faces significant challenges: the high computational cost of deep learning models and the scarcity of large-scale standardized tactile datasets. To address these limitations, this paper proposes a lightweight, sensor-agnostic computational framework accelerated by Linear Predictive Coding (LPC). Instead of relying on data-hungry end-to-end learning, our approach utilizes LPC to efficiently compress friction-induced vibration (FIV) signals into a compact spectral feature vector, which is then classified by a single-hidden-layer Fully Connected Network (FCN). Experimental results demonstrate that this framework achieves a classification accuracy of 95.2% on a custom piezoelectric testbed, exceeding benchmarks such as LSTMs and CNNs. Furthermore, the framework exhibits strong sensor-agnostic generalizability, achieving an accuracy of 97.1%–99.5% in public datasets involving barometric and capacitive sensors. Hardware implementation on a Zynq-7020 FPGA confirms its suitability for resource-constrained edge applications, consuming minimal logic resources, and achieving a sub-millisecond end-to-end latency of 0.243 ms. This work provides a practical, low-latency solution for intelligent tactile perception at the IoT edge.