Incremental learning for human action recognition in Doppler radar

The Doppler radar human action recognition methods based on deep learning theory have been developed rapidly. However, in practical applications, the offline batch training of traditional deep learning methods is difficult to adapt to the dynamic recognition needs of changing data. To solve the above problems, an incremental Doppler radar-based human action recognition class algorithm is proposed, allowing new action classes to be added gradually. Aiming at the problems of catastrophic forgetting and poor security of old class data storage, which are common in traditional incremental learning algorithms, a bias calibration method is proposed to compensate for the classifier's output bias on the old and new class data to inhibit the catastrophic forgetting effect; and a method that retains the training parameters of the generative adversarial network for generating the sample data of the old class, which avoids the risk of the leakage of the real data. Finally, the comparison experiments between the proposed algorithm and the traditional incremental learning algorithm for human action recognition are accomplished by using Doppler radar to collect the echo data of human actions, and the results verify the superior performance of the proposed method in learning new tasks while retaining the previous knowledge