Sensor-based Activities Recognition On Mobile Platform

In the past, user activities recognition was mainly achieved by computer vision.But in recent decades, with the rapid development of hardware, more and more low-cost, low-power sensor devices, data processing units and wireless communicationtechnologies are increasingly integrated into the mobile devices. From low-level datacollection to high-level applications and service delivery, pervasive computinggradually shifts from concept to implementation, which has attracted a large numberof sensor-based activities recognition. Currently, the main research of activitiesrecognition is still done on the host, so that it will limit the application of this study.If we use the mobile platform to complete the data pocessing task, it would eliminatethis limitation. The problem that newly generated is how to balance the performanceof activities recognition and its constrained data processing capabilities.In our paper, the study includes four aspects, pre-processing the raw data fromthe sensors, activities segmentation, features extraction and activities recognition,and introduces the implementation of various parts. In the activities segmentationpart, we modified the sliding window method, and use the idea of division andreorganization to extract the variable-length data of target activities from the sampleddta stream. Then, we use two algorithms as our activities recognition model, anddescribe the model training process, compare their structure and training costs.Finally, we use actual sample data, after pre-processing and features extracting,we can get the inputs of two models. In the experiments, we found that the trainingtime of HMM based activities recognition model is much less than ANN basedmodel, and the performance of HMM on time-series data is better. Meanwhile, eachactivity corresponds to a HMM model, thus it has good scalability.