Behavior Recognition Based On Different Crowd Characteristics Learning

In recent years,with the improvement of the utilization rate of smart phones,human behavior recognition technology based on smart phone accelerometer has been further developed.Compared with image based behavior recognition,behavior recognition based on accelerometer can better reflect the characteristics of human activities.The traditional behavior recognition based on acceleration sensor is mostly focused on improving the accuracy and efficiency of the classifier,neglecting the diversity of the human body features.It is limited to the use of a single general classifier model,making the accuracy of the behavior recognition to a certain bottleneck.In the data preprocessing stage,it is also simple to use filtering,windowing and other noise reduction methods,lacking a process to distinguish behavior data.In this paper,based on human behavior endpoint detection,a hybrid multiple classifier model based on human body features is adopted to conduct behavior recognition.The main work is as follows:First,an improved two parameter two threshold behavior endpoint detection algorithm is proposed.The collection environment is divided into: the smart phone is only subjected to the simple environment of the human body and the gravity of the earth's gravity and the complex environment that bears other external forces when collecting the acceleration sensor data.According to the parameters of the combined acceleration and amplitude area of the simple environment,the high threshold and low threshold value are set up to detect the behavior endpoints respectively.Two,an adaptive two parameter dual threshold behavior endpoint detection algorithm is proposed.In the stage of feature based crowd classification,the high threshold and low threshold values of each classification population are calculated.In the stage of behavior endpoint detection,the feature similarity of user and different classified population is calculated first,and the high threshold and low threshold of the user are set as the high threshold and low threshold value of the nearest classified population with characteristic similarity.Three,an improved hybrid multiple classifier model based on human characteristics is proposed.The model is divided into different groups of people,and the behavior recognition model is trained for different groups of people.The similarity between users and different classified groups is calculated.The classification population with the closest feature similarity is the group of users.In the behavior recognition stage,the user selects the behavior classifier of the group to conduct behavior recognition.Through verification experiments,it is proved that the adaptive two parameter dual threshold behavior endpoint detection algorithm can better mark the starting point and end point of "activity" behavior.The improved hybrid multiple classifier model based on human characteristics has higher recognition accuracy than other methods.