层内调节特征金字塔的密集人群姿态估计算法

A Dense Crowd Pose Estimation Algorithm for In-layer Adjustment Feature Pyramid

摘要:

针对现有姿态估计算法对密集人群检测存在漏检和误检等问题，提出一种改进YOLOv8sPose的密集人群姿态估计算法YOLOv8Pose-Dense Crowd（YOLOv8Pose-DC）。首先，设计一种集中式层内调节特征金字塔网络，采用并联方式把可变形注意力机制和CASPPF结合起来，通过自上而下的方式对金字塔网络进行全局集中调节，增加网络中全局表示的空间权重，使得改进算法能够获得全面且具有区分性的特征表示；其次，提出多尺度双检测头结构，减少计算量同时提高模型检测效率；再次，使用DySample模块，提高模型上采样效率；最后，加入上下文感知模块，提高模型全局信息关联能力，并抑制无用背景突出人物特征。实验结果表明，相较于基准模型，YOLOv8Pose-DCmAP@0.5提升3.1%，召回率提升4.2%。设计算法性能有较大提升，完全满足生产需要。

Abstract:

To address the issues of missed and false detections in dense crowd detection with existing pose estimation algorithms, an improved YOLOv8sPose algorithm for dense crowd pose estimation, named YOLOv8Pose-Dense Crowd (YOLOv8Pose-DC), is proposed. Firstly, a centralized intrinsic adjustment feature pyramid network is designed, which combines deformable attention mechanisms and CASPPF in a parallel manner. It globally focuses and adjusts the pyramid network from top to bottom, increasing the spatial weight of global representation within the network. This enables the improved algorithm to obtain comprehensive and distinctive feature representations. Secondly, a multi-scale dual detection head structure is proposed, which reduces computational complexity while enhancing model detection efficiency. Furthermore, the DySample module is utilized to improve the model""s upsampling efficiency. Lastly, a context-aware module is added to enhance the model""s ability to associate global information and suppress irrelevant background features, highlighting human characteristics. Experimental results show that, compared to the baseline model, YOLOv8Pose-DC increases mAP@0.5 by 3.1% and recall rate by 4.2%. The designed algorithm significantly improves performance and fully meets the needs of production.