Exploiting Local Temporal Characteristics via Multinomial Decomposition Algorithm for Real-Time Activity Recognition

With the current expansion of the Internet of Things (IoT), tracking human physical activity through wearable devices is becoming increasingly frequent. A human activity recognition (HAR) module is needed to localize and classify the patterns of acquired signals. Although this topic has been a talking point in the past decades, the tradeoff between efficiency, reliability, and computational complexity is still an open research challenge. In this work, we propose a novel recognition process based on online feature vector computation followed by a multinomial decomposition algorithm (MDA). Specifically, the temporal characteristics of performed activities are encoded over local segments. Afterward, a low-cost algorithm divides the multinomial classification into different stages, where the encoded patterns feed a neural network at each stage to classify the corresponding segment. The proposed approach has been evaluated on five public datasets and compared with a set of state-of-the-art methods. Experimental results show that with a few hundreds of floating-point operations per second (flops), our solution can achieve competitive performance in terms of accuracy. The effects of the sliding window length, the sampling frequency, the size of the training set, and overfitting have been studied. A prototype was developed to display the behavior of this solution in real-world conditions.