The Research Of Speaker Independent Human Activity Recognition Based On Smart Phone Multi-sensor

In recent years, the activity recognition and specific activity detection based on sensors have achieved great development. And the activity recognition based on wearable sensors research has a major position, and provides various upper application with a lot of support, such as family health, nursing injuries, etc. Wearable sensors can obtain individual sensor data in time by monitoring and then can efficiently recognize the current individual’s activities, aiming to avoid bad things from happening. Using the wearable sensor devices to be monitored will bring some unnecessary trouble to individual normal life, an d mobile sensors just can make up for the shortage because of its easy to carry, good hidden features.This paper is committed to the activity recognition research based on smart phone sensors, and HMM was adopted to realize the speaker-independent activity recognition system. Much noise problem is this article main difficulties we met. In addition, the current the HMM algorithm of local optimal problems, and on how to noise reduction and use of the HMM algorithm fusion is one of the major problems of this paper to solve.This paper analysis principle and main process of smart phone sensors ’ activity recognition in depth, using the HMM model for speaker-independent recognition, on the HMM model of the three classic problem has carried on the comprehensive understanding and analysis, aiming at Baum Welch convergence is slow and can not get the global optimal solution of the weaknesses, using the k-means algorithm to improve it.Using filtering algorithm the moving average filter and first-order low pass filter for sensor data noise filter processing and analyzing the filtering effect, then the activity recognition system platform construction and the function of the sensor data collecting module, data preprocessing module, model training module and activity recognition module are introduced and analyzed.Aiming at the shortcomings of the HMM algorithm, we proposed an HMM algorithm which is based on clustering and another one is based on discriminant analysis, and through comparing the experimental results, run them on the features extracted data set and the original data set which only retain nine features, compare and analysis the results of the experiments comprehensively.