A lightweight network model for human activity classifiction based on pre-trained mobilenetv2

Radar-based human activity classification has attracted more and more attention in the fields of anti-terrorism maintenance, post-disaster search and rescue, medical monitoring and smart homes, which not only has the advantage of detecting in harsh environments but also can protect personal privacy to the greatest extent. The combination of deep learning methods and micro-Doppler spectrograms for radar human activity recognition is becoming more and more popular, but the general network model has many parameters and long training time, making it difficult to train a better network model. Therefore, we have designed an efficient and lightweight network model, which is based on the Mobilenetv2 model pre-trained on the ImageNet dataset. The model we proposed can learn the micro-doppler information and can automatically identify tasks of different types of human movements. We use the micro-Doppler spectrogram as the training data of the network model. The model has a small number of parameters. It can train a model very quickly. At the same time, it can trace-off the accuracy and resource requirements. It can be run on the embedded in the system, it can meet the requirements of the human activity recognition system. We have shown very good performance on the experimental data provided by the official, which can not only ensure the accuracy of the model to predict human activities, but also save a lot of training time. At the same time, our experiments found that a single human actvitiy can be quickly and accurately recognized, however, the complex human behavior composed of a series of single actions still needs to continue to work