Traditional gesture recognition methods are often susceptible to interference from factors such as lighting conditions, obstructions, and inconvenient equipment carrying. This paper proposes a sensorless gesture recognition algorithm based on a multi-branch CNN-GRU framework (SGMCG). The method firstly processes Channel State Information (CSI) signals to construct a Body-coordinate Velocity Profile (BVP) dataset. Subsequently, multiple parallel CNN branches with varying kernel sizes extract spatial features of gestures. Each branch is followed by GRU networks to capture temporal dynamics. The temporal features from all branches are then fused and classified via a Softmax layer. Experimental results demonstrate that SGMCG achieves higher average recognition accuracy on both numerical and interactive gesture datasets compared to Widar3.0, GRU, and LSTM models, while also exhibiting superior robustness.