Research Of Gait Behavior Recognition Based On Event-driven Strategy

With the rapid development of sensor and wireless communication techniques, body area network has been more and more widely applied in medical, sports, entertainment and other fields. Human behavior recognition is a main hotspot in the scientific research field of body sensor network. By recognizing human behavior, computers can perceive and make sense of motivations, which provides valuable services in application scenarios of patients monitoring and athletes training, etc.Gait behavior is a common kind of movements in daily life. Recognizing gait behavior has universal significance on complicated behavior recognition in some areas. Existing gait recognition methods tend to obtain accelerometer data by slide window. And base on the data in windows, feature extraction and classifier training are finished. These methods can hardly determine the occurred time of gaits, which leads to repetitive data processing and increases the calculation amount. In addition, the existing methods can only recognize a small number of gaits without considering direction. For this reason, a gait behavior recognition method based on event-driven strategy is proposed in this paper, which can recognize 12 kinds of gaits such as walking, running, squatting, jumping, turning, going upstairs and downstairs, etc. The proposed method uses gyroscopes as main sensors, which can perceive movements of lower limbs by sampling angular velocities of legs and waist. According to the change rule of angular velocities of shanks, gait cycles are divided dynamically. And then,20 heuristic features with signification are extracted from each cycle. These features are given to a classification model for the specific gait category at last. In the process of designing the gait classification model, two classifiers are constructed respectively for the start and follow-up gaits as considering that there are certain differences in behavior between them. By experimenting four common machine learning algorithms, the best classifiers are determined.The experimental results show that the gait recognition method proposed in this article can recognize 12 kinds of gaits effectively. The average recognition accuracy of different human can be up to 98.0%. Compare to other methods, it has the characteristics of more recognizable actions, better real-time performance, higher accuracy and less calculation.