Improved 1D-CNNs for behavior recognition using wearable sensor network

Abstract

Wireless Body Sensor Network (BSNs) are wearable sensors with varying sensing, storage, computation, and transmission capabilities. When data is obtained from multiple devices, multi-sensor fusion is desirable to transform potentially erroneous sensor data into high quality fused data. By analyzing and processing the perceived physical activity data of users, they can be provided with services that may be needed. Wearable sensors transmit human acceleration information to the server through 4G network. In this way, online analysis and recognition of human behavior is realized. In this paper, we study on efficiently real-time behavior recognition algorithm using acceleration sensor. We propose a human behavior recognition method based on improved One-Dimensional Convolutional Neural Networks(1D-CNNs). According to the motion characteristics, the eigenvalues are extract which can distinguish the types of activities. At the same time, we propose a sample autonomous learning method, which aims to find the optimal sample training set and avoid over-fitting problems in traditional CNNs. In the recognition of 11 human activities, our method can reach the average accuracy of 98.7%. Compared with other behavior recognition methods in the same dataset, better classification is achieved by this method.