Active Binaural Auditory Perceptual System for a Socially Interactive Humanoid Robot

Sound Source Localization (SSL) and gaze shift to the sound source behavior is an integral part of a socially interactive humanoid robot perception system. In noisy and reverberant environments, it is non-trivial to estimate the location of a sound source and accurately shift gaze in its direction. Previous SSL algorithms are deficient in the optimum approximation of distance to audio sources and to accurately detect, interpret, and differentiate the actual sound from comparable sound sources due to challenging acoustic environments. In this article, a learning-based model is presented to achieve noiseless and reverberation-resistant sound source localization in the real-world scenarios. The proposed system utilizes a multi-layered Gaussian Cross-Correlation with Phase Transform (GCC-PHAT) signal processing technique as a baseline for a Generalized Cross Correlation Convolution Neural Network (GCC-CNN) model. The proposed model is integrated with an efficient rotation algorithm to predict and orient toward the sound source. The performance of the proposed method is compared with the state-of-art deep network-based sound source localization methods. The findings of the proposed method outperform the existing neural network-based approaches by achieving the highest accuracy of 96.21% for an active binaural auditory perceptual system.

EEG Dipole Source Localization using Deep Neural Network

Electroencephalography (EEG) inverse problem is a typical inverse problem, in which the electrical activity within the brain is reconstructed based on EEG data collected from the scalp electrodes. In this paper, the four-layer concentric head model is used for simulation firstly, four deep neural network models including a multilayer perceptron (MLP) model and three convolutional neural networks (CNNs) are adopted to solve EEG inverse problem based on equal current dipole (ECD) model. In the simulations, 100,000 samples are generated randomly, of which 60% are used for network training and 20% are used for cross-validation. Eventually, the generalization performance of the model using the optimal function is measured by the errors in the rest 20% testing set. The experimental results show that the absolute error, relative error, mean positioning error and standard deviation of the four models are extremely low. The CNN with 6 convolutional layers and 3 pooling layers (CNN-3) is the best model. Its absolute error is about 0.015, its relative error is about 0.005, and its dipole position error is 0.040±0.029 cm. Furthermore, we use CNN-3 for source localization of the real EEG data in Working Memory. The results are in accord with physiological experience. The deep neural network method in our study needs fewer calculation parameters, takes less time, and has better positioning results.

Classification of lower limb motor imagery based on iterative EEG source localization and feature fusion

Motor imagery (MI) brain–computer interface (BCI) systems have broad application prospects in rehabilitation and other fields. However, to achieve accurate and practical MI-BCI applications, there are still several critical issues, such as channel selection, electroencephalogram (EEG) feature extraction and EEG classification, needed to be better resolved. In this paper, these issues are studied for lower limb MI which is more difficult and less studied than upper limb MI. First, a novel iterative EEG source localization method is proposed for channel selection. Channels FC1, FC2, C1, C2 and Cz, instead of the commonly used traditional channel set (TCS) C3, C4 and Cz, are selected as the optimal channel set (OCS). Then, a multi-domain feature (MDF) extraction algorithm is presented to fuse single-domain features into multi-domain features. Finally, a particle swarm optimization based support vector machine (SVM) method is utilized to classify the EEG data collected by the lower limb MI experiment designed by us. The results show that the classification accuracy is 88.43%, 3.35–5.41% higher than those of using traditional SVM to classify single-domain features on the TCS, which proves that the combination of OCS and MDF can not only reduce the amount of data processing, but also retain more feature information to improve the accuracy of EEG classification.