基于深度学习的高效人体姿态估计算法的研究与实现
首发时间:2024-03-20
摘要:为了解决目前人体姿态估计算法模型由于参数量巨大,所耗费的算力资源庞大而难以在工业实践中落地的问题,本文提出了一种轻量化人体姿态估计算法模型。该模型以基于坐标分类的方法为估计手段,采用EfficientNetV2作为轻量级神经网络的骨干结构,用于对输入图像进行特征提取。为提升姿态估计的准确性,本文引入了门控注意力单元,以有效挖掘关键点的空间特征。此外,本文提出了一种混合一致性交叉熵,通过将模型的本轮次预测结果与人工标注的数据相结合,作为本轮样本数据的真实概率分布,以降低人工标注数据的误差对模型的负面影响。本文所提出的模型在COCO数据集上获得了71.7 AP的成绩,超过了许多同类轻量化模型。
For information in English, please click here
The research and implementation of an efficient human pose estimation algorithm based on deep learning
Abstract:To address the challenge of the current human pose estimation algorithm models being difficult to deploy in industrial practice due to their massive parameter size and the substantial computational resources they consume, this research proposes a lightweight human pose estimation algorithm model. The model employs a coordinate-based classification approach and utilizes EfficientNetV2 as the backbone structure of a lightweight neural network for feature extraction from input images. To enhance pose estimation accuracy, this study introduces gated attention units to effectively explore spatial features of key points. Additionally, a hybrid consistency cross-entropy is proposed, which combines the model\'s current round prediction results with manually annotated data to serve as the true probability distribution of the current round sample data, mitigating the negative impact of manual annotation data errors on the model. The proposed model in this paper achieves a performance of 71.7 AP on the COCO dataset, surpassing many similar lightweight models.
Keywords: Computer Science and Technology Human Pose Estimation Deep Learning
基金:
引用
No.****
同行评议
勘误表
基于深度学习的高效人体姿态估计算法的研究与实现
评论
全部评论