Abstract: In light of detection uncertainty in Human Activity Recognition（HAR）, the calculation method?of potential function in Markov Logic Networks（MLN） is improved. In the method, relational operators in first order logic are softened to make the binary features?extend to the interval[0, 1]; the credibility of sensor event is calculated to obtain probability of corresponding ground atom. On the basis of improved MLN, a hybrid HAR framework combined ontology is proposed and corresponding algorithm is implemented. Experimental result shows that improved MLN still has high recognition accuracy in case of ADL-E dataset contained errors.

Key words: detection uncertainty, human activity recognition, Markov Logic Networks（MLN）, event credibility, activity ontology

摘要： 针对人类活动识别中存在的检测不确定问题，改进了马尔可夫逻辑网络（MLN）中势函数的计算方法。即软化一阶逻辑中关系运算符，使特征函数的取值范围从布尔值扩展到[0，1]区间；计算传感器事件的可信度，来获取所对应闭原子为真的概率。将改进的MLN方法与本体结合，提出混合识别框架并实现了相应算法。仿真实验结果表明，在包含错误的数据集ADL-E下，改进的MLN仍能保持较高的准确率。

关键词: 检测不确定, 人类活动识别, 马尔可夫逻辑网络, 事件可信度, 活动本体