摘要
文中旨在研究基于轮廓图像空频域特征的舞蹈翻腾姿态识别模型。该模型先将待识别舞蹈视频图像实施腐蚀、膨胀、中心归一化等预处理,利用处理后图像提取舞蹈翻腾姿态能量图,通过离散余弦变换提取舞蹈翻腾姿态能量图频域特征,利用Contourlet变换提取舞蹈翻腾姿态能量图空域轮廓特征,采用特征级融合方法融合以上特征获取舞蹈轮廓图像的空频域特征向量集,再将待识别舞蹈视频序列候选姿态利用Baum-Welch算法训练为隐马尔可夫模型,利用舞蹈轮廓图像的空频域特征向量集将隐马尔可夫模型量化至观察序列,通过前向后向算法获取观察序列姿态概率,观察序列概率值最大的隐马尔可夫模型对应姿态即为所需识别舞蹈翻腾姿态。实验结果表明,该模型可较好地提取具有空频域特征的舞蹈轮廓图像,有效识别舞蹈视频中舞蹈翻腾姿态,且识别100帧有阴影舞蹈视频图像中舞蹈翻腾姿态识别准确率高于96%。
The dance somersault attitude recognition model based on contour image spatial frequency domain features is researched in this paper. In this model,the dance video images to be identified are preprocessed with corrosion,expansion and center normalization,with which the dance posture energy diagrams are extracted. The frequency domain features of the dance posture energy diagrams are extracted with the discrete cosine transform,the spatial frequency domain features of the dance posture energy diagrams are extracted with the Contourlet transform,and then the above features are fused by means of the featurelevel fusion method to obtain the spatial frequency domain feature vector sets of the dance contour images. Any more,the candidate attitudes of the dance video sequence to be identified are trained for the hidden Markov model by means of the Baum-Welch algorithm,which is quantified to the observation sequence with the spatial frequency domain feature vector sets of the dance contour images. The attitude probability of observation sequence is obtained by means of the forward-backward algorithm,and the corresponding attitude of the hidden Markov model with the highest probability value of the observation sequence are the dancing somersault attitude needed to be identified. Experimental results show that this model can well extract dance contour images with spatial frequency domain features,effectively identify dance somersault postures in dance video,and the accuracy rate of dance somersault posture recognition in shadowed dance video images with 100 frames is higher than 96%.
作者
耿君
GENG Jun(Shandong Normal University,Jinan 250014,China;Shandong University of Finance and Economics,Jinan 250014,China)
出处
《现代电子技术》
北大核心
2019年第24期146-149,153,共5页
Modern Electronics Technique
基金
国家自然科学基金项目(81460582)~~
关键词
舞蹈翻腾姿态
姿态识别
轮廓图像
空频域特征
模型训练
对比验证
dance somersault attitude
attitude identification
contour image
model training
comparison experiment
