Human Gait Prediction And Recognition Based On Deep Learning And Inertial Data

With the arrival of the aging population and the number of people with physical disabilities caused by traffic accidents is increasing.These people with mobility difficulties need external help to help them carry out sports activities and monitor their physical condition.The emergence of wearable devices,to a certain extent,brings solutions to these problems.As a kind of wearable devices,exoskeleton is widely used in medical,assistance,and other fields.However,the current wearable exoskeleton still has some problems.For example,the wearable exoskeleton is not suitable for human movement behavior,which is easy to cause a sense of restraint to the wearer;according to the gait characteristics of different groups,the recognition effect is not ideal,which can not achieve the recognition accuracy required to complete the auxiliary movement.In order to overcome the shortcomings of gait recognition and prediction in human lower limbs,this paper proposes a gait recognition and prediction method based on inertial sensors and deep learning.Firstly,aiming at the gait data set,a lower limb motion inertial data acquisition platform is built by using Noitom motion capture system,which can accurately and effectively capture the human lower limb inertial data in motion.All the actions in the self-built data set are acquired by inertial sensors located in the thighs,legs and feet of the left and right legs.Each sensor collects six channels of inertial data,and all data is 36 channels in total.The movements collected in this paper mainly include standing,walking,jogging,up and down stairs,up and down the slope.Secondly,aiming at gait recognition and prediction,this thesis mainly studies gait recognition and prediction based on inertial data and the problem of response lag caused by exoskeleton calculation delay.Considering that the gait data is a kind of time series data,this paper proposes the gait neural network(GNN)model based on the temporal convolutional network.The model is mainly composed of the intermediate module,the target vector coded module and the action recognition and prediction module,which can fully extract the historical information in the gait data series.The results of experiments show that GNN model achieves the best performance in all indicators of gait prediction task,and also meets the application requirements in gait recognition task,and the comprehensive performance is improved compared with other methods.Finally,in order to achieve the better generalization performance of the model for gait data of different groups,a gait recognition method based on orthogonalization is proposed.Based on the idea of cosine similarity,the inner product of gait recognition feature vector and identity recognition feature vector is calculated;then the result of the inner product is back propagated as a part of the loss function,which makes the value of the inner product close to zero,so as to achieve the orthogonalization of two feature vectors approximately.This method can improve the recognition effect of gait behavior of different groups,and effectively improve the generalization ability of the model.The experimental results show that for the data set with less samples,there are less hidden features in the data,and the algorithm model proposed in this thesis improves significantly compared with other traditional methods;when the number of samples increases,the algorithm model proposed in this paper also achieves the comprehensive optimal recognition effect.