Efficient human pose estimation in complex coal mining scenes via Keypoint Partitioning Adaptive Convolution
Jin Wu, Huaping Zhou, Xiangrui Meng, Tao Wu · Alexandria Engineering Journal · 2026
Human pose estimation (HPE) is crucial for underground mining safety, but it suffers from uneven brightness, occlusions from dense equipment, complex backgrounds, and limited computational resources. To address these challenges, we propose a novel Keypoint-Adaptive Convolutional Network (KAnet) for accurate miner pose estimation. KAnet integrates our newly proposed content-adaptive convolution method called Keypoint Partitioning Adaptive Convolution (KAconv), which adaptively partitions feature maps based on semantic similarity and generates region-specific dynamic filters. This design enables the model to handle complex and variable spatial information distribution effectively. Additionally, we introduce an Attention-Based Cross-Layer Feature Fusion (ACFF) module to enhance multi-scale feature fusion and improve robustness against occlusion and illumination variations. To further optimize model efficiency, we present the Pruning-guided Adaptive Filtering Knowledge Distillation (PAF-KD), which leverages channel importance ranking for efficient model compression while preserving essential feature representations. We validate the effectiveness of KAnet using the newly developed Miner-Pose dataset, a large-scale dataset of miner poses in coal mines. Experimental results demonstrate that KAnet outperforms current state-of-the-art methods in both accuracy and robustness in complex mining scenarios.