基于时空模式的复杂行为识别方法研究

Sensing and recognizing human behavior under complex environment has been heatedly discussed and widely researched in the area of pervasive computing. Focusing on key issues such as representation, modeling and recognizing complex individual behavior, this project studies the techniques to layering, representation and recognizing complex behavior from the aspect of behavior perception and recognition. It also researches on a multi-sensor perception based complex individual behavior recognition system, utilizing theories and methods from disciplines of pattern recognition, machine learning, data mining and human-machine interaction. It first studies self-adaptive window partition, feature extraction and optimization, as well weighted feature fusion-based atomic motion recognition method, through raw data acquisition of mobile objects in real-world by deploying multi-typed sensors. Based on the analysis of behavioral characteristics and patterns of mobile objects, the project then studies the spatio-temporal correlation among atomic activities, and constructs a hierarchal model to describe and recognize individual complex behavior under spatio-temporal environment. Finally, through mining frequent spatio-temporal patterns among complex atomic activities, the project researches on various types of complex behavior recognition methods such as sequential, concurrent and intersecting method. It also establishes a complex behavior recognition system under spatio-temporal environment, which provides a new perspective for future research on sensor-based activity recognition.

感知和识别复杂环境中人的行为是普适计算研究的热点和难点课题之一。本项目针对个体复杂行为描述、建模和识别等关键问题，运用模式识别、机器学习和数据挖掘等理论和方法，研究基于多传感器感知的复杂行为识别方法体系。首先，基于人的行为的差异性和多模性特点，研究多模式行为的自适应检测与数据划分方法；并通过分析行为模式、关键特征子集与分类算法三者之间的关联关系及特征之间的交互性，研究基于特征子集区分度的行为识别特征优选方法，进而研究基于时空特性和加权特征融合的原子动作识别方法。其次，通过分析人的行为特性和规律，研究原子动作间的时空相关性，构建基于时空模式的个体复杂行为描述与识别分层模型。在此基础上，通过挖掘构成复杂行为的原子动作间的频繁时空模式作为中间层特征，研究顺序、并发和交叉等类型复杂行为的识别方法，建立基于时空模式的复杂行为识别方法体系，为基于传感器的行为识别方法研究和应用提供新的思路和理论依据。

人体行为识别主要是指对人体的动作类别、行为模式进行分析和识别，并对其加以描述，具有广阔的应用前景和可观的经济价值。本项目针对人体行为描述、建模和识别等关键问题，运用模式识别、机器学习和数据挖掘等理论和方法进行研究。针对行为识别过程中同类动作的周期性、不同类动作的交替性以及动作切换时的过渡动作等多模式识别问题，提出了一种基于马氏距离的行为数据分割方法，实现了多行为模式过程中基本动作与过渡动作数据的准确分割。针对传统特征优选方法大多对特征之间相关性考虑不足的问题，以分类准确性为准则，提出了一种基于特征相关度的二进制粒子群特征优选方法。由于单一类型或单一位置的传感器不能全面准确地识别日常行为活动问题，因此提出了一种基于加权线性判别分析的多传感器数据融合识别方法。考虑到手机的位置及在真实环境中的使用，如静坐、行走、上楼梯、下楼梯、乘坐公交等，基于集成分类器的识别方法被用来提高不同场景下的识别精度。本项目针对行为识别中数据分割、特征优选、数据融合和行为识别方法的探索，为相关研究和应用提供了一种新的思路，具有较好的理论和实际意义。