行为识别领域的研究已持续多年，而随着科技的不断创新，其研究方向也大范围扩展。以基于传感器的行为识别为基础服务，通过对个人行为状态的识别以及行为状态持续时间的统计，以监控、提醒、报警等方式可以解决时下热点问题中的“空巢老人”无人看管期间身体出现状况无人发现问题、独守儿童缺乏监护等问题。数据的准确性决定着行为识别的结果，利用配备多项传感器的智能手机可以采集手机使用者处于各种行为状态时的传感器数据。从传感器数据中提取相关特征构成特征集合，利用采集的样本训练分类器并用分类器对所有行为进行判决是行为识别的实现方式之一。 从传感器数据所构成的数据样本中提取相关特征，利用不同的特征组合训练分类器，最终行为识别的分类准确率不尽相同；对所有相关特征进行不同组合，选择维数低、判决准确率高的特征组合是解决行为识别的关键步骤。本文提出了一种基于遗传算法的行为识别特征选择算法，将智能手机传感器所采集的数据特征进行预处理和特征提取，利用遗传算法从所有特征构成的集合总集中选择行为识别正确率最高的特征集合。 基于加速度传感器的行为识别是目前较为热门的解决行为识别的方法，尤其是智能手机的廉价和便携，使加速度数据更容易采集。因为加速度数据是矢量，其受手机放置的位置和方向的影响很大，即使是静止不动的行为状态，由于手机摆放的不同而产生的加速度数据差异也会很大。因此，本文针对该问题提出了多次分类的行为识别方法，使用合成加速度的方式来进行一次分类，判定行为状态，若状态为爬楼，则利用加速度信号的数值分量对爬楼行为进行二次分类，确定行为模式为上楼或者下楼，通过对比实验确认了其行为识别正确率确实较单次分类有了明显提升。 通过设计实验最终验证，本文所提出的基于遗传算法的行为识别特征选择方法可以在特征总集中选择维数低、判决准确率高的特征组合；通过本文提出的多次分类方法对行为识别进行判决可以提升行为的识别率。

The field of behavioral recognition has been studied for many years. with the continuous innovation of science and technology，the scope of research has also expanded greatly. Using sensor based behavioral identification as basic service，problems, such as "Empty Nester" and children who lack of guardianship, can be solved through identification of individual behavior and statistics of behavior status by monitoring, giving reminders and alarming.The outcome of behavioral identification is highly depending on the data source. User’s behavioral data in different status can be monitored through various types of sensors equipped on a smart phone.Extracting the relevant features from the data collected to form a feature set, Up on which the classifier can be trained by making different decisions so as to realize behavioral identification. The accuracy of behavioral recognition is different by the classifier which is trained through different feature sets which are formed by using a variety of combinations.This paper presents a behavioral recognition feature selection algorithm based on genetic algorithms，which abstracts the data features collected by the smart phone and performing preptreatments and extraction up on them, finally picking the feature set which results in a highest recognition rate by using genetic algorithm. Behavior recognition based on accelerometer is a popular method to solve behavior recognition, especially smart phone's low cost and portability, so that acceleration data can be easily collected. The acceleration data is a vector，its direction is controlled by positions and directions of cell phone. Even if the current status of human behavior is standing still，the acceleration data would be quite divergent under different orientation of the phone. This paper has brought up a solution which applies a multiple behavioral classification solely against such issue, which performs a primary classification according to the aggregate triaxial acceleration data to justify the behavior of current status, if the status of climbing stairs is recognized then the secondary classification would be executed against the real part of the acceleration vector so as to identify whether it is up or down. Therefore, it is obvious that multiple classification method is more effective compare to the solution only does single classification in respect of the rate of successful identification. Through the final verification of designed experiments, the proposed method is able to pick feature combinations with low dimensionality and high accuracy. And the accuracy of behavior identification can also be greatly improved through multiple classification method.