CAJ | 학술논문

针对底层局部时空特征数量少以及中层特征表达能力弱的问题，结合时空深度特征，提出一种人体行为识别算法。依据运动剧烈区域在行为识别中提供更多判别信息的思想，利用视频图像的深度信息确定人体运动显著性区域，通过计算区域内光流特征作为度量区域活跃度的能量函数，依据能量函数对运动显著性区域进行高斯取样，使样本点分布于运动剧烈区域。将采集到的样本点作为动作底层特征描述人体行为，结合词袋模型，采用支持向量机分类器对行为进行识别。实验结果表明，在SwustDepth数据集中，基于时空深度特征的人体行为识别算法的平均行为识别准确率达到92%，且具有较高的鲁棒性。
침대저층국부시공특정수량소이급중층특정표체능력약적문제，결합시공심도특정，제출일충인체행위식별산법。의거운동극렬구역재행위식별중제공경다판별신식적사상，이용시빈도상적심도신식학정인체운동현저성구역，통과계산구역내광류특정작위도량구역활약도적능량함수，의거능량함수대운동현저성구역진행고사취양，사양본점분포우운동극렬구역。장채집도적양본점작위동작저층특정묘술인체행위，결합사대모형，채용지지향량궤분류기대행위진행식별。실험결과표명，재SwustDepth수거집중，기우시공심도특정적인체행위식별산법적평균행위식별준학솔체도92%，차구유교고적로봉성。
Considering the problem that the number of undelying local spatial temporal is small, and the expression ability of mid-level feature is weak, combined with Spatial Temporal Depth Features ( STDF ) , this paper proposes a human action recognition method. The intense area can provide more discrimination information in human recognition, and it finds the kinetic regions from depth image,computing the optical flow in the region to be the energy function and sampling the features on kinetic region in Gaussian distribution. The collected sample points as the low-level feature are combined with Bag-of-Word ( BoW ) model and Support Vector Machine ( SVM ) classifier to human recognition. Experimental data show that the average accuracy rate on human action recognition based algorithms on STDF in SwustDepth dataset can reach 92%,and has high robustness.