Online semi-supervised learning classifier for radar posture recognition, and classification method of classifier

Abstract

The invention discloses an online semi-supervised learning classifier for radar posture recognition and a classification method of the classifier. The classifier comprises: a feature buffer module, which is used for performing sample caching under a preset replacement strategy; and an extreme random forest module, which has classification capability after offline pre-training, wherein newly arrived label-free posture samples are endowed with pseudo labels and are subjected to screening, the indexes, the pseudo labels and the feature vectors of the label-free posture samples are stored into the feature buffer module, the indexes into are stored corresponding leaf nodes of the extreme random forest module, label counters in the leaf nodes are updated according to the pseudo labels, and when the leaf nodes meet splitting conditions, samples of corresponding batches are extracted from the feature buffer module according to the stored index vectors, and a node splitting algorithm of an extreme random forest is called for splitting. Therefore, the problem of high cost of offline collection of training samples and the problem of low generalization performance of an offline supervised machine learning algorithm on users who do not provide training samples are solved.

Description

technical field [0001] This application relates to the technical field of target radar recognition, in particular to an online semi-supervised learning classifier and its classification method for radar gesture recognition. Background technique [0002] Radar-based gesture recognition is a non-contact human-computer interaction method, which is widely used in wearable devices, smart home, virtual reality and other fields due to its advantages of low power consumption, low privacy leakage and low dependence on light conditions. Broad application space. [0003] The workflow of radar-based gesture recognition can be roughly divided into: the transmission of gesture detection signals, the reception and processing of echo signals, feature extraction and classification. The front-end based on coherent pulse signal and the signal processing method based on range-Doppler are widely used because they can effectively describe the position and velocity information of the motion postu...

AI Technical Summary

Problems solved by technology

[0006] This application provides an online semi-supervised learning classifier and its classification method for radar pose recognition, so that the classifier can use the unlabeled user pose samples collected successively during the system to adjust its structure and parameters online, and overcome the existing offline The problems of high training sample collection cost and low cross-user recognition rate common in supervised learning algorithms improve the accuracy of user gesture recognition and solve the problem of high cost of offline collection of training samples and offline supervised machine learning algorithms are not provided. The problem of low generalization performance on users with over-trained samples

Images