基于LPN改进的轻量化姿态估计方法ULPN
Improved Lightweight Attitude Estimation Method ULPN Based on LPN
高彦彦 1任好盼 2危德健3
作者信息
- 1. 32738部队 郑州 450053
- 2. 北京理工大学计算机学院 北京 100081
- 3. 字节跳动有限公司 北京 100080
- 折叠
摘要
在人体姿态估计过程中,往往采用复杂的网络结构实现较好的预测性能,但是模型实际推理速度较慢.针对该问题,论文提出了新的轻量化姿态估计网络ULPN,采用轻量化姿态估计网络LPN作为基础,利用改进的轻量化网络和新型的轻量化瓶颈模块提升模型推理效率.论文提出了基于Ghost卷积模块和注意力机制模块GCB的新型轻量化瓶颈模块,减少冗余的特征图同时对全局建模.利用连续卷积层代替池化层能够保存更多有效信息,采用分组卷积降低模型的计算负载,同时基于轻量化瓶颈模块提出了轻量化姿态估计网络ULPN.该算法在相似的预测精度下,能够有效降低模型的计算负载,较好地进行实时人体姿态估计.
Abstract
In the process of human pose estimation,complex network structures are often adopted to achieve better prediction performance,but the actual inference speed of the model is slow.In response to this problem,the paper proposes a new lightweight human pose estimation network ULPN,which uses the lightweight human pose estimation network LPN as the basis,and uses an im-proved lightweight network and a new lightweight bottleneck block to improve model inference efficiency.The paper proposes a new lightweight bottleneck block based on the Ghost convolution module and the attention mechanism block GCB to reduce redundant feature maps and simultaneously model the global.Using a continuous convolutional layer instead of a pooling layer can save more ef-fective information,and grouped convolution is used to reduce the computational load of the model.At the same time,a lightweight human pose estimation network ULPN is proposed based on the lightweight bottleneck block.With similar prediction accuracy,the algorithm can effectively reduce the computational load of the model,and better perform real-time human pose estimation.
关键词
深度学习/姿态估计/模型轻量化/推理速度/预测精度Key words
deep learning/pose estimation/lightweight model/inference speed/predict accuracy引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据