Abstract: To address the issues of poor generalization and low recognition accuracy in existing millimeter-wave radar-based human pose recognition methods, we propose a multi-channel feature selection dense network (MCF-SE-DenseNet). This model integrates multi-channel feature (MCF) fusion to improve the DenseNet architecture. Leveraging the antenna array configuration, radar echoes are processed to extract micro-Doppler information, elevation information and azimuth information. These multi-channel features are fused into a 3D matrix to spatially characterize human posture with higher fidelity. The Squeeze-and-Excitation (SE) module, an attention mechanism, is embedded into the densely connected network. The fused 3D matrix is converted into feature maps as input for the enhanced DenseNet, which dynamically amplifies the weights of critical features, thereby significantly improving pose recognition accuracy. Experimental results demonstrate that the combined use of multi-channel feature fusion and the SE module increases recognition accuracy by more over 6.2%, effectively enhancing model performance. The final recognition accuracy of the MCF-SE-DenseNet model exceeds 98%.