Smartphone sensors-based human activity recognition using feature selection and deep decision fusion

Human activity recognition (HAR) with smartphone sensors is a significant research direction in human-cyber-physical systems. Aiming at the problem of feature redundancy and low recognition accuracy of HAR, this paper presents a novel system architecture comprising three parts: feature selection based on an oppositional and chaos particle swarm optimization (OCPSO) algorithm, multi-input one-dimensional convolutional neural network (MI-1D-CNN) utilizing time-domain and frequency-domain signals, and deep decision fusion (DDF) combining D-S evidence theory and entropy. The proposed architecture is evaluated on the UCI HAR and WIDSM datasets. The results highlight that OCPSO performs better than particle swarm optimization (PSO) in feature selection, convergence speed, and recognition accuracy. Moreover, it is shown that for the MI-1D-CNN classifier, the frequency-domain signals (95.96%) perform better than time-domain signals (95.66%). In addition, this paper investigates the impact of the convolution layers, feature maps, filter sizes, and decision fusion methods on recognition accuracy. The results demonstrate that the DDF method (97.81%) outperforms single-layer decision fusion in improving the recognition accuracy on the UCI HAR dataset.