train

wireless signal space, Doppler Frequency Shift (DFS) is affected by gesture orientation, position and other factors, which leads to the difference of spectral mode characteristics of similar gestures. Previous cross-domain gesture recognition systems need to pre-label factors unrelated to gestures, or use multiple links to enhance gesture information. The acquisition of cumbersome prerequisites and the increase of links limit the performance and universality of cross-domain gesture recognition systems. In this paper, we propose Wi-Ellipse, a cross-domain dynamic modeling system for hand gestures from implicit conditions contained in DFS, which does not require any prerequisite conditions. Our system not only achieves exciting classification results in hand gesture and human orientation estimation, but also achieves considerable accuracy in a wide range of human activity tasks. The key of Wi-Ellipse is to model the frequency of the wireless space, infer millisecond regional motion information through the spatial frequency model, design a network model similar to Recurrent Neural Network (RNN) to estimate the motion direction, and dynamically model the gesture through the region and direction information. Finally, a CNN+RNN hybrid deep neural network was designed to classify the modeling results. Under the test of various data sets and classification tasks, Wi-Ellipse is a lightweight, fast, accurate and universal human activity recognition system. The accuracy of single cross-domain and cross cross-domain gesture recognition reaches 93.8%~96.8% and 96.6% respectively, and the accuracy of cross cross-domain gesture orientation recognition reaches 92%. Under the universal link, the cross-domain