Accurate human locomotion and localization using hybrid features and deep networks on smartphones

Human locomotion and localization recognition are of critical importance in the development of intelligent systems with the capability of monitoring, analyzing, and responding to human behavior in the real world. This research proposes a robust yet computationally efficient approach by exploiting the sensors of a smartphone, such as accelerometers, gyroscopes, and Global Positioning System (GPS). The proposed approach improves the accuracy of localization and locomotion recognition. The proposed approach is computationally efficient and yet robust. The approach begins with the preprocessing of the sensor signals by using the Savitzky-Golay filter for noise reduction. The signals are then subjected to overlap-add windowing. The signals are then subjected to a comprehensive feature extraction process. To ensure the discriminability of the features and eliminate redundant features, Luca-measure Fuzzy Entropy (LFE) optimization technique has been used. The technique helps in identifying non-linear relationships effectively. The features are then subjected to a Graph Convolutional Network (GCN) for classification. The proposed approach has been tested with the ExtraSensory and Sussex-Huawei Locomotion (SHL) datasets. The results show localization accuracy of 92.57%, 89.83%, locomotion recognition of 93.18%, and 91.75%, respectively.