基于多模联合分类的校园暴力与言语欺凌识别算法研究

School bullying is a common social problem which affects teenagers both physically and mentally, so bullying detecting techniques are essential for school bullying prevention. However, existing anti-bullying methods mainly focus on passive methods of feeding-back. This project first detects physical violence and verbal bullying events actively by pattern recognition theory. By analyzing the differences between violent activities and daily-life ones, we build a mix-coordinate to extract suitable action features and design a composite-mode fuzzy classification algorithm. The proposed algorithm can detect physical violence among children of different ages and distinguish them from daily-life activities. We build mathematical models to compare the acoustic feature relationship between bullying voices and daily-life ones, and design an iterative increasing and decreasing feature selection algorithm by both theoretical analysis and simulation. This algorithm can select the best acoustic feature set for verbal bullying recognition. We also develop a feature fitting algorithm to analyze the components of the mixed emotion. Then we combine the mixed speech recognition and emotion analyzing algorithms to distinguish verbal bullying from daily-life speeches. Finally we use a speech-based mental stress recognition algorithm to assist the activity recognition and the emotion recognition, and a joint judgment integrating the classification confidence and matching similarity is made to detect school bullying events. This project first uses theoretical models to analyze the relationship among features, and involves action-voice combined features to recognize bullying events in complex environments, thus has important scientific significance.

校园霸凌是一种常见的社会现象，伤害青少年的身心健康，研究霸凌检测技术对及时制止校园霸凌意义重大。目前对校园霸凌的防治都是被动的反馈方式，本课题首次用模式识别理论对身体暴力和言语欺凌进行主动检测。分析霸凌动作与日常动作行为模式上的差异，建立混合坐标系提取动作特征，设计复合模式模糊分类算法，在不同年龄对象群体中检测出身体暴力，并与日常行为区分开。建立数学模型研究霸凌语境与日常语境的声学特征关联，从理论分析和仿真实验两方面共同设计迭代增减量特征优选算法选取最佳声学特征组合，同时用特征模型拟合解析多重情感成分，结合语境识别和情感解析区分霸凌语境和日常语境。最后用基于复杂语境声学特征的精神压力识别算法辅助动作识别和情感识别，结合分类置信度和匹配相似度做联合判决，检测校园霸凌事件。本课题首次用理论模型分析特征关联，以动作语音联合特征识别复杂环境中的霸凌事件，具有重要的科学意义。

校园霸凌是一种常见的社会现象，危害青少年的身心健康。校园霸凌有多种表现形式，其中身体暴力对青少年的伤害最大，言语欺凌发生最为频繁。传统的霸凌防治手段是人工的，即旁观者向监护人报告，然而旁观者可能因害怕报复而不敢报告，霸凌得不到及时制止。为此，本项目利用模式识别技术，提出能够主动检测校园霸凌事件，并在检测出后自动向监护人告警的算法。本项目包含三项研究内容，其中前两项是出于受害者自身角度进行检测，而第三项是出于监护人的角度进行检测：（1）基于运动传感器的身体暴力检测算法。利用运动传感器采集佩戴者多个部位动作数据，根据暴力动作和日常动作行为模式上的差异，提取时域和频域特征。提出一种改进的Relief-F特征选择算法，在筛除无用特征的同时降低选用特征间的冗余度，然后通过LDA算法对特征进一步降维。设计DT-RBF算法对动作进行识别，并提出一种改进的D-S算法，对不同部位识别结果进行融合，最终对暴力动作识别率达90%，对日常动作识别率达95%。（2）基于语音情感的言语欺凌检测。利用麦克风采集佩戴者周围的语境信息，首先进行语音预处理，然后提取语境中的MFCC特征和一阶、二阶差分特征，利用LDA算法对所提取的声学特征进行降维。设计BPNN分类算法对霸凌语境和日常语境进行区分，因识别目标是混合语境，平均识别率仅为73.6%。当言语欺凌同时伴随有暴力动作时，结合暴力动作分类，对言语欺凌的识别率可达90.9%。（3）基于图像的霸凌场景检测。利用校园内监控摄像头采集指定区域图像，利用KNN算法提取前景目标并通过形态学处理提取准确的运动人体。提出一种外接矩形框整合方法，消除非观测对象，然后提取观测对象的形态学特征、外接矩形框特征和光流特征。利用Relief-F和Wrapper算法筛除无用特征。设计DT-SVM算法对霸凌场景和日常场景进行分类，平均识别率达97.4%。最后，本项目已开发出基于Android手机的校园霸凌检测演示平台，因手机性能受限，不能使用高复杂度算法，且只有单一传感器，暴力动作识别率仅为86.7%。