Deep Learning Based Abnormal Gait Classification System Study with Heterogeneous Sensor Network
Yin, Shubao (2020-01-20)
Deep Learning Based Abnormal Gait Classification System Study with Heterogeneous Sensor Network
Yin, Shubao
(20.01.2020)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2020062345512
https://urn.fi/URN:NBN:fi-fe2020062345512
Tiivistelmä
Gait is one of the important biological characteristics of the human body. Abnormal gait is mostly related to the lesion site and has been demonstrated to play a guiding role in clinical research such as medical diagnosis and disease prevention. In order to promote the research of automatic gait pattern recognition, this paper introduces the research status of abnormal gait recognition and systems analysis of the common gait recognition technologies. Based on this, two gait information extraction methods, sensor-based and vision-based, are studied, including wearable system design and deep neural network-based algorithm design.
In the sensor-based study, we proposed a lower limb data acquisition system. The experiment was designed to collect acceleration signals and sEMG signals under normal and pathological gaits. Specifically, wearable hardware-based on MSP430 and upper computer software based on Labview is designed. The hardware system consists of EMG foot ring, high-precision IMU and pressure-sensitive intelligent insole.
Data of 15 healthy persons and 15 hemiplegic patients during walking were collected. The classification of gait was carried out based on sEMG and the average accuracy rate can reach 92.8% for CNN. For IMU signals five kinds of abnormal gait are trained based on three models: BPNN, LSTM, and CNN. The experimental results show that the system combined with the neural network can classify different pathological gaits well, and the average accuracy rate of the six-classifications task can reach 93%.
In vision-based research, by using human keypoint detection technology, we obtain the precise location of the key points through the fusion of thermal mapping and offset, thus extracts the space-time information of the key points. However, the results show that even the state-of-the-art is not good enough for replacing IMU in gait analysis and classification. The good news is the rhythm wave can be observed within 2 m, which proves that the temporal and spatial information of the key points extracted is highly correlated with the acceleration information collected by IMU, which paved the way for the visual-based abnormal gait classification algorithm. 步态指人走路时表现出来的姿态,是人体重要生物特征之一。异常步态多与病变部位有关,作为反映人体健康状况和行为能力的重要特征,其被论证在医疗诊断、疾病预防等临床研究中具有指导作用。为了促进步态模式自动识别的研究,本文介绍了异常步态识别的研究现状,系统地分析了常见步态识别技术以及算法,以此为基础研究了基于传感器与基于视觉两种步态信息提取方法,内容包括可穿戴系统设计与基于深度神经网络的算法设计。
在基于传感器的研究中,本工作开发了下肢步态信息采集系统,并利用该信息采集系统设计实验,采集正常与不同病理步态下的加速度信号与肌电信号,搭建深度神经网络完成分类任务。具体的,在系统搭建部分设计了基于MSP430的可穿戴硬件设备以及基于Labview的上位机软件,该硬件系统由肌电脚环,高精度IMU以及压感智能鞋垫组成,该上位机软件接收、解包蓝牙数据并计算出步频步长等常用步态参数。
在基于运动信号与基于表面肌电的研究中,采集了15名健康人与15名偏瘫病人的步态数据,并针对表面肌电信号训练卷积神经网络进行帕金森步态的识别与分类,平均准确率可达92.8%。针对运动信号训练了反向传播神经网络,LSTM以及卷积神经网络三种模型进行五种异常步态的分类任务。实验结果表明,本工作中步态信息采集系统结合神经网络模型,可以很好地对不同病理步态进行分类,六分类平均正确率可达93%。
在基于视觉的研究中,本文利用人体关键点检测技术,首先检测出图片中的一个或多个人,接着对边界框做图像分割,接着采用全卷积resnet对每一个边界框中的人物的主要关节点做热力图并分析偏移量,最后通过热力图与偏移的融合得到关键点的精确定位。通过该算法提取了不同步态下姿态关键点时空信息,为基于视觉的步态分析系统提供了基础条件。但实验结果表明目前最高准确率的人体关键点检测算法不足以替代IMU实现步态分析与分类。但在2m之内可以观察到节律信息,证明了所提取的关键点时空信息与IMU采集的加速度信息呈现较高相关度,为基于视觉的异常步态分类算法铺平了道路。
In the sensor-based study, we proposed a lower limb data acquisition system. The experiment was designed to collect acceleration signals and sEMG signals under normal and pathological gaits. Specifically, wearable hardware-based on MSP430 and upper computer software based on Labview is designed. The hardware system consists of EMG foot ring, high-precision IMU and pressure-sensitive intelligent insole.
Data of 15 healthy persons and 15 hemiplegic patients during walking were collected. The classification of gait was carried out based on sEMG and the average accuracy rate can reach 92.8% for CNN. For IMU signals five kinds of abnormal gait are trained based on three models: BPNN, LSTM, and CNN. The experimental results show that the system combined with the neural network can classify different pathological gaits well, and the average accuracy rate of the six-classifications task can reach 93%.
In vision-based research, by using human keypoint detection technology, we obtain the precise location of the key points through the fusion of thermal mapping and offset, thus extracts the space-time information of the key points. However, the results show that even the state-of-the-art is not good enough for replacing IMU in gait analysis and classification. The good news is the rhythm wave can be observed within 2 m, which proves that the temporal and spatial information of the key points extracted is highly correlated with the acceleration information collected by IMU, which paved the way for the visual-based abnormal gait classification algorithm.
在基于传感器的研究中,本工作开发了下肢步态信息采集系统,并利用该信息采集系统设计实验,采集正常与不同病理步态下的加速度信号与肌电信号,搭建深度神经网络完成分类任务。具体的,在系统搭建部分设计了基于MSP430的可穿戴硬件设备以及基于Labview的上位机软件,该硬件系统由肌电脚环,高精度IMU以及压感智能鞋垫组成,该上位机软件接收、解包蓝牙数据并计算出步频步长等常用步态参数。
在基于运动信号与基于表面肌电的研究中,采集了15名健康人与15名偏瘫病人的步态数据,并针对表面肌电信号训练卷积神经网络进行帕金森步态的识别与分类,平均准确率可达92.8%。针对运动信号训练了反向传播神经网络,LSTM以及卷积神经网络三种模型进行五种异常步态的分类任务。实验结果表明,本工作中步态信息采集系统结合神经网络模型,可以很好地对不同病理步态进行分类,六分类平均正确率可达93%。
在基于视觉的研究中,本文利用人体关键点检测技术,首先检测出图片中的一个或多个人,接着对边界框做图像分割,接着采用全卷积resnet对每一个边界框中的人物的主要关节点做热力图并分析偏移量,最后通过热力图与偏移的融合得到关键点的精确定位。通过该算法提取了不同步态下姿态关键点时空信息,为基于视觉的步态分析系统提供了基础条件。但实验结果表明目前最高准确率的人体关键点检测算法不足以替代IMU实现步态分析与分类。但在2m之内可以观察到节律信息,证明了所提取的关键点时空信息与IMU采集的加速度信息呈现较高相关度,为基于视觉的异常步态分类算法铺平了道路。