Multi-Stream Head Pose Estimation Algorithm Based on Enhanced Feature Extraction

To address the challenges of poor real-time performance and low detection accuracy in existing head pose estimation algorithms under complex scenarios, this paper proposes a Multi-Stream Head Pose Estimation algorithm based on Feature Enhanced Extraction for Multiple Stream Network (FEEM-Net). First, a three-branch parallel structure is designed with different activation functions and pooling methods to enhance the diversity of feature expression at the same level through multi-channel feature extraction. Second, the Convolutional Block Attention Module (CBAM) is introduced after the pooling layer to focus the head region features by using the channel and spatial attention mechanism to effectively suppress the background interference. Finally, a bottleneck residual module based on asymmetric convolution is proposed to enhance the modeling ability of multi-scale and multi-directional information, and improve the information flow transfer efficiency by residual connection. Experimental results show that the proposed algorithm reduces the MAE to 4.63 and 4.06 on the AFLW2000 and BIWI datasets respectively.