DCNN-DGML融合深度神经网络步态识别算法  

Gait Recognition Algorithm Based on DCNN-DGML Fusion Deep Neural Network

作　　者：王欢 肖秦琨[1] WANG Huan;XIAO Qin-kun(School of Electronic and Information Engineering,Xi’an Technological University,Xi’an 710021,China)

机构地区：[1]西安工业大学电子信息工程学院,西安710021

出　　处：《自动化与仪表》2021年第11期84-89,95,共7页Automation & Instrumentation

摘　　要：针对骨架步态识别任务中步态信息时空关联性弱和易受复杂背景干扰的问题,该文提出一种基于深度卷积神经网络和动态门记忆学习(DCNN-DGML)的骨架步态识别算法。首先根据步态运动规律建立端到端的步态识别模型;其次为了实现多层特征语义信息的融合,该模型中设计了多层特征融合机制改进GoogleNet;同时,在LSTM的基础上设计了一种动态门记忆学习结构来记录时序信息,以提升特征的鲁棒性。为了评估提出的DCNNDGML步态识别系统,在自建数据库上进行了实验研究,识别精度达到97%,结果明显优于传统步态识别方法。In skeleton-based gait recognition task,the spatio-temporal correlation of gait information is weak and the data is vulnerable to complex background interference.In order to solve the above-mentioned problems,the paper proposes a skeleton gait recognition algorithm based on deep convolutional neural network and dynamic gate memory learning(DCNN-DGML).Firstly,an end-to-end gait recognition model is established according to the gait motion discipline;secondly,in order to achieve the fusion of multi-layer feature semantic information,the multi-layer feature fusion mechanism is designed in the model to improve GoogleNet;meanwhile,a dynamic gate memory learning structure is designed on the basis of LSTM to record timing information to improve the robustness of features.Based on the self-built database,the extensive experiments are carried out to evaluate the performance of the proposed DCNN-DGML gait recognition system,and the recognition accuracy is 97%.The results are obviously better than traditional gait recognition methods.

关 键 词：步态识别 骨架序列 深度卷积神经网络 长短期记忆网络 动态门记忆学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]