Nonlinear oscillator reservoirs for real time visual tracking
Abstract
This work introduces a physical reservoir computing (PRC) framework for real-time prediction of object trajectories within camera-captured visual frames, utilizing nonlinear Duffing oscillator arrays as dynamic reservoirs. We study two configurations: an uncoupled array, where each oscillator encodes a single input channel, and a coupled array that enables interactions across channels. Motion signals derived from a calibrated webcam and ArUco tracking are normalized and mapped to bifurcation-sensitive operating regions to elicit rich yet stable dynamics. The reservoir states are sampled and concatenated, and only a linear readout is trained via ridge regression, enabling efficient learning with minimal computational overhead. We evaluate next-step and multi-step forecasts under noisy and cluttered scenes and observe strong short-horizon accuracy with resilience to abrupt trajectory changes. The Duffing-based PRC offers a compact, low-computation alternative to deep sequence models for embedded applications in robotics, surveillance, and autonomous systems.
Publication Title
Communications in Nonlinear Science and Numerical Simulation
Recommended Citation
Gupta, A., & Agarwal, V. (2026). Nonlinear oscillator reservoirs for real time visual tracking. Communications in Nonlinear Science and Numerical Simulation, 163 https://doi.org/10.1016/j.cnsns.2026.110795
