Human Activity Recognition Based On AHRS In Intelligent Space

The research on human activity recognition is the important prerequisite and key component for human behavior understanding and friendly human-computer interaction. With the development of information and intelligence, the robots and intelligent devices can identify the person’s actions and understanding behavior intentions stably and accurately, and how to make them work for human more friendly has become a hot spot argument in recent years.With the rapid development of MEMS technology, wearable sensors show their advantages in terms of human posture recognition gradually. Wearable sensors are no longer affected by the lighting, background and other external circumstances. Moreover, this method overcomes shortcoming of camera limited area. Therefore, wearable sensors are used by more and more researchers to identify action and understand behavior.In this paper, aim at the human activity recognition, a wearable sensor module named attitude and heading reference system (AHRS) is designed. It is mainly composed of CPU, gyroscope, accelerometer and magnetometer, etc. Extended Kalman Filter (EKF) was used for multi-sensor data fusion, which calculates the stable and accurate attitude angle, angular velocity and acceleration. Based on the AHRS module, a platform of human motion capture has been set up to provide basic data for gesture recognition.The gestures referred in this paper are defined as some special actions whose template is invariable, and the type of activity is numbered. So, Dynamic Time Warping (DTW) is suitable for the gesture recognition. A new method that Endpoint Detection (ED) based on Dynamic Time Warping algorithm is prestented in this paper, which can classify the gesture activity using the characteristics (three-axis attitude Angle and the motion acceleration).As we all konwn, the foundation of behavior understanding is the recognition of human basic activity. Human basic activities described in this paper refer to the basic unit of the human daily behavior, such as walking, sitting down, squatting down and bending over. Hidden Markov Model (HMM) is established for each basic action. The activity characteristics captured by sensors are regard as the observation sequence to calculate the probability, and identify the basic action. As a result of these basic activities have the context relationship, Hierarchical Hidden Markov Model (HHMM) is set up to narrow the scope of activities. It not only improved the recognition rate, but also reduced the computational cost.Human complex activities are proposed in the paper, such as drinking water, making a phone call, sweeping and so on. In order to meet the requirements of the complex actvity characteristics integrity, five AHRS modules are installed on the key parts of the trunk and limbs to extract action feature, and multi-sensor data are fused to achieve complex activities recognition.The human fall detection device with high recognition rate has great significance to reducing the damage of the elderly or patients caused by falling. In this paper, posture angle and three axis acceleration data provided by AHRS module fixed on the human waist are used as the input for the human fall detection system, and a novel method based on Back Propagation (BP) neural network is proposed for fall detection. The fall activity and daily behavior can be carefully distinguished accurately used by this method.Large numbers of experimental results demonstrate that the hardware platform designed is stable and reliable, and the ED-DTW algorithm, HHMM algorithm, CHMM algorithm and BP neural network algorithm have the higher recognition rate and good robustness.