The Study On Activity Recognition Method Based On Intelligent Terminal

Action recognition is a hot research topic in the computer field,which has achieved novel applications in different fields,such as health care,security and entertainment.With the rapid development of information age,the development of micro electromechanical system(MEMS)technology and the appearance of micro sensors,intelligent terminals have gradually become a powerful tool for studying behavior recognition.The method of behavior recognition based on intelligent terminal has the characteristics of portability,high robustness and no impact on the normal life of users.Because the sensor in the intelligent terminal can detect the user's motion data,the intelligent terminal can be used to improve the recognition accuracy.However,the recognition rate of this method is still relatively low,especially the recognition rate of similar activities is lower,and the computational complexity is high.All these problems lead to the low practicability of the recognition system.Therefore,in order to further solve the above problems,this thesis has done relevant research and put forward corresponding solutions.In view of the low recognition rate of similar activities in current activity recognition,this thesis proposes a recognition system for similar activities in smartphones.The three-dimensional acceleration data are converted into five-dimensional vectors through formulas to improve the validity of the data.Then,the optimal feature set is obtained by feature extraction and feature selection algorithm.Finally,Multi-layer Perception classifier is used for classification and recognition.The experimental results on the open online database show that the average recognition rate of model proposed in this thesis is 99.2%,which is about 3% higher than the existing results.The accuracy of similar activities increased by about 5%.Aiming at the problem that the accuracy of fall recognition is not high and the false alarm rate is high.Based on the threshold detection method,a hybrid method combining threshold detection with machine learning detection is proposed in this thesis.First,a pre-judgment is made on the activity.If the condition is met,the pre-judgment is a fall,and the behavior pre-judged as a fall is further classified by machine learning.On the contrary,it is prejudged as a daily activity and continues to be monitored.Experimental results show that this method not only improves the accuracy of recognition,reduces the false positive rate,but also reduces the complexity of the algorithm.