A Class Incremental Temporal-Spatial Model Based on Wireless Sensor Networks for Activity Recognition

With the development of Internet of Things (IoT), all kinds of sensors appear everywhere in our lives. Sensor-based Human Activity recognition (HAR) as the most common application of IoT system, has attracted attentions from both academia and industry due to the popularity of IoT. In most real-life scenarios, HAR is dynamic. For instance, in the process of classifying the activities, a new class is often encountered. To effectively recognize new activities, we propose a three-stage class incremental temporal-spatial model based on wireless sensor networks for activity recognition. In the first stage, when the new class arrives, rather than using the traditional method to train all the old class data, we design a method based on the recognition score to select part of the old class data as a sample for training, which represents the data distribution of the old class data. In the second stage, we analyze the temporal and spatial characteristics of activity data in combination with physical knowledge, the model extract temporal-spatial features, which further enhances the recognition effect. In the third stage, we avoid the phenomenon of 'catastrophic forgetting' of a network by finetuning distillation loss and classification loss. To demonstrate the effectiveness of our proposed model, we conducted comparative experiments using different models on different public datasets. The results showed our proposed model can not only recognize new classes, but also maintain the recognition ability of old classes.