Category-Extensible Human Activity Recognition Based on Doppler Radar by Few-Shot Learning

Radar-based human activity recognition (HAR) has been applied in many fields such as human-computer interaction, smart surveillance, and health assessment. With the development of deep learning, many deep-learning models have been proposed in radar-based HAR to achieve high classification accuracy. Inevitably, new human activity classes appear continuously, and this requires the deep-learning models to efficiently learn novel categories from only a few data samples while at the same time maintaining high accuracy on the initial human activities on which they were trained. Therefore, this article proposes a category-extensible HAR model based on few-shot learning. The classification weight of the new human activity category is quickly generated through the feature vector extracted by the trained feature extractor and the classification weight of the most similar original category. To solve the mismatching of weight value intervals, we also employ the cosine similarity-based classifier. Furthermore, we adopted the large margin in the softmax cross-entropy (LMSC) loss function to improve the model's performance and depthwise separable convolution to reduce the computation of the model. The experimental results show that the model can quickly adapt to the new human activity classification task and did not sacrifice the performance on category classification of the original activities. With relatively low computational complexity, our method achieves 81.99% of accuracy when each extended category has only one sample to generate the weights and 91.60% when each of them has only five samples in the task of a three-way.