Specific Human Following by Residual-Bi-LSTM-Based Distributed Module UWB Network and Residual-RNN-Based Finite-Time Control

Abstract

To implement the specific human following (SHF) task in a global GPS-denied environment (e.g., airport, hotel, hospital, inventory area, office, library), a distributed module ultra wideband network (DM-UWBN) is deliberated on residual-bi-long short-term memory (RBLSTM) model. It possesses these advantages: maintain its gradient signal for more effective learning, reduce the overfitting weight, and improve localization accuracy. Besides the dynamic localizations of omnidirectional service robot (ODSR) and specific human, a residual-recurrent neural network -based finite-time control (RRNN-FTC) based on the relative degree of ODSR's output is designed. It has a dominant advantage in the converging weight learning compensation of aggregately dynamic uncertainties. Based on these valuable characteristics, outstanding tracking performance with reducing average power or fluctuation of control input in contrast to previous research is achieved. Due to the multiprocessing times of RRNN-FTC (e.g., 5 ms) and RBLSTM-DM-UWBN (e.g., 150 ms), the SHF is more arduous. Finally, the SHF of ODSR by the RRNN-FTC in the global GPS-denied environment with two pillars and a mat on the reference path validates the supremacy of the proposed approach.