CAS-HDPA-Yolov11: a dual attention enhanced network for monocular human height estimation

Abstract

Monocular human height estimation is of great value in intelligent security inspection and surveillance applications, but existing methods often suffer from insufficient object detection accuracy and the high computational cost of attention mechanisms. This paper proposes a dual-attention enhanced network based on improved YOLOv11, named CASHDPA- YOLOv11. First, the Convolutional Additive Self-attention (CAS) module replaces C2PSA in the backbone to achieve lightweight global context modeling. Second, the Hierarchical Dual-Path Attention (HDPA) module is inserted into the neck network to refine head-foot region features through dual paths of global average pooling and local max pooling. On the VOC2007 pedestrian detection task, our model achieves an mAP50 of 83.8% and a precision of 1.00. On a 10-subject test in a simulated security inspection scenario, the mean absolute error (MAE) of height estimation is reduced from 8.88 cm (baseline) to 2.95 cm, the relative error from 5.03% to 1.67%, and the success rate (error < 5 cm) increases from 20% to 70%, while inference speed remains at 0.107 seconds per image. Our method demonstrates the feasibility of a lightweight dual-attention synergistic mechanism that significantly improves vertical boundary localization at the cost of a slight decrease in general detection accuracy, providing a new approach for fine-grained measurement in structured scenarios with a fixed monocular camera.