Activity Recognition Based On Ubiquitous Computing

Wearable activity recognition is one of the hot issues in the field of ubiquitous computing.It uses wearable sensors to recognize human behavior.Compared with video-based behavior recognition,wearable behavior recognition can overcome the interference of dark environment or partial occlusion,and can monitor human behavior for a long time.Despite its high application and research value,this task faces two challenges.First of all,the traditional artificial feature engineering is a time-consuming and heuristic method,and can not make full use of the original sensor data information.Secondly,for the behavior recognition problem under multi-sensor conditions,the spatial correlation between sensors provides important information for behavior recognition.How to model the spatial interaction relationship between sensors is also one of the problems that need to be solved.Current research proves that deep learning can learn powerful feature representations from raw data.In order to solve the above difficulties and challenges,the main contributions and innovations of this paper are as follows:1.For the behavior recognition under the condition of single sensor,we design a multi-view learning model HAR-Net,and propose to use a multi-scale separation convolution network to extract deep features,And prior knowledge is introduced through artificial feature engineering.2.For behavior recognition under multi-sensor conditions,a novel framework GraphConvLSTM is proposed,which integrates graph convolution,temporal convolution and LSTM,the model can learn local temporal interaction and long-term temporal dependency from raw sensor data,and can model the spatial relationship between different sensors.We evaluated the proposed model on real single-sensor and multi-sensor data sets,and the experimental results show that our model’s performance in activity recognition exceeds the existing methods.