Research On Attention Mechanism For Wearable Human Activity Recognition

With the development of wearable computing and deep learning technology,human activity recognition(HAR)based on wearable devices are more and more applied to medical,security,entertainment,military,and other areas.HAR is drawing widespread academic attention.The traditional HAR method based on computer vision is subject to computer computing power,strong privacy intrusion,and easy to be restricted by multiple factors such as occlusion,light,and background.Wearable computing is a computing mode that "wears" embedded devices on the body.The wearable device can continuously perceive human activity in real-time and isn’t affected by environmental factors.Moreover,wearable technology provides the benefits of compact size,portability,and affordability.It is currently a hot topic in research area.At this stage,the activity recognition methods based on wearable devices are mainly faced with the following problems: 1.Long-range dependency on temporal sequence data is ignored.Long-range correlation information is ignored in the HAR method based on CNN.Although the HAR method based on LSTM networks can establish long sequence dependencies,it will repeatedly model unrelated signal features.2.The large-scale receptive field causes feature compression.The HAR method based on Convolutional Block Attention Module(CBAM)uses the large-scale receptive field to extract the temporal correlation feature and the sensor channel correlation feature,resulting in the compression of low-semantic sensor features.3.The self-attention based HAR method loses local context information.The HAR method based on the self-attention mechanism loses the context information of the sensor signal when establishing the association between the local feature and the global feature.In regard to these three problems,This thesis propose the following three methods:1.This thesis propose a HAR method based on positional attention mechanismIn view of the loss of long range temporal dependency by convolutional networks and cross channel correlation,existing methods generally use LSTM network and CNN network to extract long-range temporal correlation and cross-sensor channel correlation.However,due to the uncorrelated signal(noise)in the sensor data,LSTM will model the noise repetitively,resulting in the reduction of the classification accuracy of the activity.This thesis propose a HAR method based on the positional attention(PA)mechanism in this thesis.This method divides the signal feature into two 1D features,extracts the long sequence temporal dependency and cross-sensor channel correlation features on the two dimensions by using the 1D global pooling operation,and embeds the attention features of the two dimensions into the CNN network respectively to enhance the network’s ability to extract the signal feature,thus increasing the accuracy of the network’s classification of activity.The HAR method based on the positional attention mechanism proposed in this thesis achieves 96.58%,97.16%,78.84%,95.40%,and 99.64% recognition accuracy on UCI-HAR,PAMAP2,Uni Mib-SHAR,DSADS,and MHEALTH datasets.2.This thesis propose a HAR method based on multi-scale channel attention mechanismTo solve the problem of feature compression caused by large-scale receptive field,existing methods have used large-scale receptive field to convolution signal features to extract temporal and cross-channel correlation of high semantic.However,it cause the feature compression for low semantic signal features,lower network resolution of signals,decreasing the classification’s accuracy.In this thesis,a HAR method based on a multi-scale channel attention(MSCA)mechanism is proposed.We use multiple receptive fields to extract signal features in parallel.Then,we combine the multi-scale feature maps,and use a cross-feature channel extraction block to establish a cross-channel association for multi-scale feature maps.The experimental results on four HAR data sets of UCIHAR,DSADS,PAMAP2,and Uni MibSHAR show that this method performs well,and the optimal recognition accuracy reaches97.20%,95.57%,97.18%,and 79.39%.3.This thesis propose a HAR method based on contextual attention mechanismTo solve the problem of losing local context information in HAR methods based on the self-attention mechanism,existing methods have established the association between Key(single point signal feature)and Query(global signal feature)through the self-attention mechanism.But,the activity signal is temporal data,and existing methods would lost the context information around Key.This thesis proposes a HAR method based on the contextual attention mechanism.This method uses the local context information of the signal features around the Key to guide the learning of the global attention information.It connects the local context information with the global information residual to prevent the activation function from losing a large amount of feature information.The experimental results on five public HAR datasets show that the classification accuracy on UCIHAR,PAMAP2,Uni Mib-SHAR,DSADS,and MHEALTH data sets is 96.99%,97.80%,80.35%,95.81%,and 99.19%,compared with the activity recognition method based on the self-attention mechanism.Finally,based on the above research results,this thesis implements online activity recognition prototype systems.This thesis collect activity data in the natural environment,deploy the data training depth model to the server,and finally prove the model’s effectiveness through online prediction.