This letter presents lower limb human motion detection using a surface electromyogram (sEMG) with a top and slope (TAS) feature extraction algorithm. Lower limb human motion detection using sEMG signal is generally divided into gait subphase detection, locomotion mode recognition, and mode change detection. Existing feature extraction algorithms using sEMG signal have several innate problems in recognizing lower limb human motion detection. With respect to time-domain features, their values may be analogous because two different gait subphases and locomotion mode may have similar muscle activity pattern. Therefore, it is not easy to select the proper feature set of sEMG signals. The TAS feature extraction algorithm reflects better timing characteristics of sEMG signals than the existing time-domain feature extraction algorithm. Therefore, it can provide high accuracy in lower limb human motion detection. Experimental results show that the average detection accuracy values of the proposed method in terms of the gait subphase detection, locomotion mode recognition, and mode change detection are increased by 8%, 5%, and 4% better than those of feature values of the Willison amplitude, respectively.

To apply emotion recognition and classification technology to the field of human-robot interaction, it is necessary to implement fast data processing and model weight reduction. This paper proposes a new photoplethysmogram (PPG) and galvanic skin response (GSR) signals-based labeling method using Asian multimodal data, a real-time emotion classification method, a 1d convolutional neural network autoencoder model, and a lightweight model obtained using knowledge distillation. In addition, the model performance was verified using the public DEAP dataset and the Asian multi-modal dataset ‘MERTI-Apps’. For emotion classification, bio-signal data were window-sliced in 1-pulse units, and the label was reset to reflect the characteristics of the PPG and GSR signals. Simple data pre-processing, such as the prevention of loss and waveform duplication, was performed without using handcrafted features. The experiment showed that the accuracy of the proposed model using MERTI-Apps was 79.18% and 74.84% in the case of arousal and valence, respectively, for 3-class criteria, and the accuracy of the proposed model using DEAP was 81.33% and 80.25% in the case of arousal and valence, respectively, for 2-class criteria. The accuracy of the lightweight model was 77.87% and 73.49% in the case of arousal and valence, respectively, for 3-class criteria and its calculation time was reduced by more than 80% compared to the proposed 1d convolutional autoencoder model. We also confirmed that the proposed model improved computational time and accuracy compared to previous studies using MERTI-Apps and the lightweight model used in limited hardware environments enabled fast computation and real-time emotion classification.

This study focuses on driver-behavior identification and its application to finding embedded solutions in a connected car environment. We present a lightweight, end-to-end deep-learning framework for performing driver-behavior identification using in-vehicle controller area network (CAN-BUS) sensor data. The proposed method outperforms the state-of-the-art driver-behavior profiling models. Particularly, it exhibits significantly reduced computations (i.e., reduced numbers both of floating-point operations and parameters), more efficient memory usage (compact model size), and less inference time. The proposed architecture features depth-wise convolution, along with augmented recurrent neural networks (long short-term memory or gated recurrent unit), for time-series classification. The minimum time-step length (window size) required in the proposed method is significantly lower than that required by recent algorithms. We compared our results with compressed versions of existing models by applying efficient channel pruning on several layers of current models. Furthermore, our network can adapt to new classes using sparse-learning techniques, that is, by freezing relatively strong nodes at the fully connected layer for the existing classes and improving the weaker nodes by retraining them using data regarding the new classes. We successfully deploy the proposed method in a container environment using NVIDIA Docker in an embedded system (Xavier, TX2, and Nano) and comprehensively evaluate it with regard to numerous performance metrics.

This paper presents a gait sub-phase detection and prediction approach using surface electromyogram (sEMG) signals, pressure sensors, and the knee angle for a lower-limb power-assist robot. Pattern recognition and machine learning models using sEMG signals have several inherent problems for gait sub-phase detection. These problems are due to recognition delay, lack of consideration for the unique characteristics of sEMG signals based on the subject, and meaningless features. To solve these problems, we propose a new labeling technique based on the heel and toe, a muscle and feature selection, a user-adaptive classifier using a weighted voting technique to achieve gait sub-phase detection, and a gait sub-phase prediction technique using interpolation. Experimental results show that the average accuracies of the proposed labeling, the muscle and feature selection, and the user-adaptive classifier using weighted voting are 7%, 12%, and 17% better, respectively, than the existing methods using physical sensors. Results also show that the average prediction time of the proposed method is 80% faster than the existing methods.