Human Motion Recognition Based On Machine Learing

Human motion recognition is becoming a research upsurge, which aims at understanding human behavior, and plays an increasingly important role in a number of applications, such as health care, smart home and drunk driving detection. The traditional way of human action recognition is by fixing the external sensors in some parts of the human body, using the information collected by the sensors to analyze human actions. This method is simple and effective, but it brings a lot of trouble to everyone's daily life. Wearable sensor is more humane than the traditional external sensors, but due to the high price, the implementation is not very common. With the rapid development of science and technology, the intelligent mobile phone sales increased year by year, the built-in sensor types are more and more abundant, and the use of them is very high frequency. So we study human behavior recognition technology based on the intelligent mobile phone, and we not only proposed sliding window, coordinate transformation which is a new method of processing data, based on the study of common action recognition on the innovation of complex action recognition algorithm with posture transformation as a supplement. A pose transition is a transient action that describes the transition from a static action to another. Because their frequency is very low and the duration is shorter than that of other actions, most of the human motion recognition methods ignore the pose shift. However, it is not possible to ignore the other requirements of the human motion recognition system, because they are more significant than others. In this paper, we propose a three layer model to recognize the human action. In experiment, we used Iphone6 to collect the basic movements, such as jumping, walking, running, upstairs and downstairs.In the recognition of basic action, this paper first analyzes and studies the concept of coordinate transformation in the inertial navigation system. Then, a coordinate transformation formula is proposed. Finally,the specific test and analysis are carried out. After preprocessing, the original time slice sequence data can be divided into discrete and easily classified training samples by the appropriate size of the sliding window.In the feature extraction phase, we want to analyze the signal from the time - frequency angle, and we need to use the appropriate wavelet basis function. Finally, based on the comparison of several classification algorithms, we choose the support vector machine as the classifier training samples. Next, in order to further improve the performance of the algorithm, combined with the idea of AdaBoost to improve the training algorithm AdaBoost SVM: SVM as a weak classifier iteration,respectively, to training data, if a sample is wrong, then increase its weight in the next iteration, if the correct classification will reduce its weight.In the gesture recognition, because the posture transformation duration is shorter than the ordinary movement, there is a certain difference between the data acquisition method and the basic movement.After getting the complete motion information, we put forward a kind of three layer model to classify the action, which can effectively identify the action of the gesture.At the end, the paper puts forward some ideas about the deficiencies in the research and the improvement of the paper.