Continuous Human Activity Recognition in IoT Environments With BAOA and AFSG-TPD GANs

We propose a method for recognizing continuous indoor daily human activities among continuous indoor Internet of Things (IoT) smart environments using millimeter wave radar. Focusing on the following problems: 1) transition errors due to random transitions between actions during continuous action recognition and 2) the time-consuming and labor-intensive factors on radar data acquisition for continuous human actions make it difficult to consider all action sequences, and the network suffers from catastrophic degradation of the recognition performance when faced with completely new human action sequences. A bounding adaptive optimization algorithm based on interlacing error (BAOA) and a generative adversarial network based on adaptive feature selection generator and temporal patch discriminator (AFSG-TPD GAN) are proposed. BAOA is used to accurately segment the existing action sequences to obtain a single action dataset of random duration, synthesize the data used to train the AFSG-TPD GAN, generate new action sequences, and train the recognition network to improve generalization performance. After the comparison test, BAOA increases the average accuracy by 3.91% compared to the state-of-the-art (SOTA) method. Meanwhile, the network trained with the data generated by the AFSG-TPD GAN overcomes the problem of catastrophic degradation of the recognition performance when confronted with brand new human action sequences in real-world tests, and the average accuracy is improved from 65.01% to 94.85%.