基于多特征融合卷积的步态识别算法研究
GAIT RECOGNITION ALGORITHM BASED ON MULTI-FEATURE FUSION CONVOLUTION
杨鹏 1应娜 1李怡菲1
作者信息
- 1. 杭州电子科技大学通信工程学院 浙江杭州 310018
- 折叠
摘要
针对GaitSet算法中主干网络学习能力和分类能力较弱,提出基于多特征融合卷积网络的步态识别算法(MFFC-GaitSet).算法通过多特征融合卷积重建GaitSet网络增强网络学习能力,同时对三元组损失函数进行平滑优化;利用形态学处理对步态轮廓图进行修补.算法在Casia-B数据集上进行验证,步态识别精度达到85.811％,提高2.6％;模型权重仅增加6％.算法可以有效减少复杂环境对步态识别的负面影响,实现复杂环境下高精度的步态识别.实验结果表明,方法能够实现较为精确的步态识别,并具有较佳的鲁棒性和泛化能力.
Abstract
Aimed at the weak learning and classification ability of the backbone network in the GaitSet algorithm,the gait recognition algorithm based on the multi-feature fusion convolution(MFFC-GaitSet)is proposed.The algorithm reconstructed the GaitSet network by multi-feature fusion convolution to enhance the network learning ability,and smoothed and optimized the ternary loss function.The gait contour map was repaired by morphological processing.The algorithm was validated on the Casia-B dataset and achieved a gait recognition accuracy of 85.811％,with the increase of 2.6％.The model weight was increased by only 6％.The algorithm could effectively reduce the negative influence of complex environment on gait recognition and achieve high-precision gait recognition in complex environment.The experimental results show that the method can achieve more accurate gait recognition with better robustness and generalization ability.
关键词
步态识别/多特征融合/形态学处理/三元组平滑优化/Casia-B数据集Key words
Gait recognition/Multi-feature fusion/Morphological processing/Triplet smoothing optimization/Casia-B dataset引用本文复制引用
基金项目
浙江省自然科学基金项目(LY16F010013)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据