Inertial Sensor-based Motion Sensing Activity Recognition And Analysis

Body moving activity recognition and analysis is a technology used for predicting and recognizing the users' status by analyzing the information about users,which can be used in human-computer interaction,kinematic analysis,behavior understanding,indoor navigation,health care and the elder guardianship,etc.After more than ten years of development,there are mainly two kinds of methods for body moving activity recognition and analysis,the first one is based on video device,which analyzes the RGB or RGBD image information obtained from video,the second one is based on inertial sensor,by acquiring the real-time signal from wearable sensors to analyze the body moving activity.Compare with the video-based methods,the motion sensor-based methods have the advantages of smaller size,lower power consumption,and are not affected by light condition and target moving.Recently,with the rapid development of mobile devices,the smart phones and smart watches mainly based on iOS and Android platform have already integrated gyroscope,accelerometer,the heart rate meter and magnetometer and all kinds of sensors,together with the growing computing performance of smart phones and smart watches,the body moving activity recognition and analysis tasks are completely competent.On the premise of hardware support,in this thesis,we proposed an activity recognition method based on convolution neural network,by modifying the convolution filter size to achieve the purpose of adaptation based on inertial sensor data.This method can demonstrate better recognition performance than the traditional method without additional feature engineering,the test accuracy of which reaching 93.8%.On the basis of this method,the characteristics of the deep residual network are further added to the network to add the identity mapping,so that the test accuracy of the network can reach 95.9%.Upon completion of the above theoretical research,in this thesis,we implemented deep residual network based real-time recognition algorithm on the Android Wear,and realized the forward algorithm by means of C++ and compiled into a static link library.Then the recognition algorithm is called on Android Wear with JNI.After successfully using the static link library of the deep residual network on Android Wear,we completed the Android Wear based exercise recognition on the smartwatch.In the end,we also studied the data generation based on generative adversarial network,and applied to kinds of body moving recognition by adding generated samples to the original training dataset for data augmentation,in order to improve the classifier's generalization ability.Experiments show that the data generation method is suitable for a variety of body moving recognition applications for data augmentation,and we find that when the number of generated samples and the number of original training data samples are comparable,the network performance is best.