Research On Human Motion Recognition Algorithm Based On Deep Neural Network

The increasing popularity of mobile intelligent devices provides researchers with a series of new research directions.Human activity recognition(HAR)based on wearable sensor data has become an important research area due to its wide application in motion tracking,health monitoring and smart home.The traditional machine learning methods for wearable sensor-based human activity recognition tasks usually require manual feature extraction,and the deep neural network that can automatically extract human activity data features is becoming a new research focus.At present,DeepConvLSTM,which is combined of convolutional neural network(CNN)and long short-term memory(LSTM)recurrent neural network,has better recognition accuracy than other recognition methods.However,the training of this model requires a large amount of computing resources.In addition,most popular neural networks are designed manually,which requires the designer to have certain professional knowledge and spend a lot of time and computing resources to build the optimal architecture for different datasets.(1)Data preprocessing and sequence segmentation of the original sensor data:the data preprocessing methods include normalization,standardization and missing value filling of the original sensor signal;Sequence segmentation is the use of fixed length of the sliding window in a certain step to segment the original sequence data into an instance of equal growth.(2)To solve the difficulty of training neural networks with long short-term memory recurrent unit,the paper proposes a fusion model based on convolutional neural network and gated recurrent unit(GRU),and the performance on three public data sets(ACT data set,UCI data set and OPPORTUNITY data set)is compared with convolutional neural network and DeepConvLSTM.Our experimental results show that the recognition accuracy of the model on three public data sets is higher than that of convolutional neural network and is close to DeepConvLSTM,but the convergence speed of the model is faster than that of DeepConvLSTM.(3)Instead of manual choosing a suitable topology,we will let the progressive neural network architecture search algorithm design the optimal topology in order to maximize the classification F1 score.The algorithm uses a sequential model-based optimization(SMBO)strategy,in which we search the structure space in order of increasing complexity,while learning a surrogate function to guide the search through the structure space.The structure we found in this way achieve state of the art classification accuracy on the OPPORTUNITY dataset,which indicates that this method is useful and can improve previously manually-designed architectures.