IMU Sensing–Based Hopfield Neuromorphic Computing for Human Activity Recognition

Aiming at the self-association feature of the Hopfield neural network, we can reduce the need for extensive sensor training samples during human behavior recognition. For a training algorithm to obtain a general activity feature template with only one time data preprocessing, this work proposes a data preprocessing framework that is suitable for neuromorphic computing. Based on the preprocessing method of the construction matrix and feature extraction, we achieved simplification and improvement in the classification of output of the Hopfield neuromorphic algorithm. We assigned different samples to neurons by constructing a feature matrix, which changed the weights of different categories to classify sensor data. Meanwhile, the preprocessing realizes the sensor data fusion process, which helps improve the classification accuracy and avoids falling into the local optimal value caused by single sensor data. Experimental results show that the framework has high classification accuracy with necessary robustness. Using the proposed method, the classification and recognition accuracy of the Hopfield neuromorphic algorithm on the three classes of human activities is 96.3%. Compared with traditional machine learning algorithms, the proposed framework only requires learning samples once to get the feature matrix for human activities, complementing the limited sample databases while improving the classification accuracy.