ACSFNet: adaptive channel–spatial fusion for efficient 2D human pose estimation toward motion analysis

Efficient two-dimensional human pose estimation is important for visual perception systems operating under limited computational and memory resources. Although LiteHRNet provides a lightweight high-resolution backbone, its compact channel operations may restrict cross-channel interaction and spatial feature selection, while conventional heatmap-based prediction introduces additional memory and decoding overhead. To address these limitations, this study proposes ACSFNet, an Adaptive Channel–Spatial Fusion Network for efficient 2D human pose estimation. First, an Adaptive Channel Transformation (ACT) module is developed by combining grouped convolution, pointwise convolution, conditional channel weighting, and input-dependent branch fusion within a shuffle-style structure. This design strengthens local and cross-channel feature interaction without introducing a heavy self-attention architecture. Second, a Spatial Dynamic Gate (SDG) module is introduced to estimate feature importance from normalized responses and perform gated cross-group reconstruction, thereby attenuating redundant activations while retaining potentially useful information. Third, an existing SimCC-based coordinate classification head is integrated into ACSFNet to replace the conventional two-dimensional heatmap prediction head and reduce prediction overhead. Experiments on the MPII validation set and the COCO 2017 benchmark show that ACSFNet-18 and ACSFNet-30 improve PCKh@0.5 by 0.6 and 0.9 percentage points, respectively, compared with their LiteHRNet baselines. On COCO, ACSFNet-30 achieves 68.0 AP with approximately 3.0 million parameters. These results indicate that the proposed channel–spatial modifications provide a favorable balance between localization accuracy and computational efficiency. ACSFNet may serve as an efficient image-based pose perception component for future resource-constrained visual analysis systems.