基于融合特征的人体动作识别

针对复杂的人体动作识别率较低以及单一特征的不足的问题,提出基于Haar小波变换的改进轮廓特征以及多种特征融合的动作识别方法。在第一组实验中对小波变换后的单轮廓特征采用动态时间规整(DTW)算法进行动作序列匹配;第二组实验中提取每帧深度图像的时空兴趣点并进行匹配,在轮廓特征基础上融合加速稳健特征(SURF)和光流直方图(HOF)运动特征作为局部特征;第三组实验在第二组实验的基础上加入骨骼特征。后两组实验均采用线性支持向量机(SVM)对提取的特征进行训练和分类,并在公开数据集MSR Action3D上进行验证。实验结果验证了改进后轮廓特征的有效性,两组融合特征的动作识别方法在同类算法中具有较高的识别率。

Human action recognition based on fusion features

Aiming at the problem of low recognition rate and the lacks of single feature of complex human actions,an improved contour feature based on Haar wavelet transform and action recognition method with multiple feature fusion were proposed.In the first group of experiments,dynamic time warping (DTW) algorithm was utilized to match template for motion sequences of the single contour features after wavelet transform.In the second experiment,the spatio-temporal interest points of each depth image were extracted and matched,and the speeded up robust features (SURF) and the histograms of oriented optical flow (HOF) motion were fused on the basis of the contour feature.In the third experiment,the skeletal features were added.The linear support vector machine (SVM) was used to train and classify the extracted features in the latter two groups of experiments,the experiments were performed on the open MSR Action3D dataset.Results show that the effectiveness of the improved contour feature is verified and the two methods of action recognition based on fusion features show higher recognition rate compared with similar algorithms.