| In the era of large data,the field of non-traditional data is rapidly expanding,and the field of sensor-based activity recognition has also changed dramatically.It is not easy to mine valuable information from massive raw data,and the time consumpution and recognition accuracy of traditional classification algorithm are challenged.This has spawned the development of many new technologies and multidisciplinary integration,such as the use of deep convolution neural networks for sensor-based activity recognition.In this paper,we study the deep convolution neural network for multi-dimensional sensor activity identification,and experiment results on multiple data sets show the proposed DCNN method achieves good performance.First,traditional SVM algorithms and DCNN are used for sensor-based activity recognition.The SVM algorithm and the depth convolution neural network are simulated on several data sets,and the results shows that the deep convolution neural network is more suitable for the active recognition in the actual scene than the SVM algorithm.Secondly,the principle,concept,overall architecture,sparse connection and weight sharing thought of convolution neural network are described in detail.Then,the convolution neural network based on the fusion feature is introduced.Compared to reference algorithms,the convolution neural network based on the fusion feature performs better.The accuracy of the experiment on opportunity dataset was 88.0%,while accuracy of the experiment on Skoda dataset was 95.2%.Finally,convolution and long-short-term memory recurrent neural networks are used for activity recognition.Based on the general framework proposed by convolution and long-short-term memory recurrent units,the feature can be extracted the former,and the time information can be captured by the latter.Experiments show that this framework can optimize the training model and further reduce the number of parameters to be trained in the case of recognition accuracy. |
PDF Full Text Request