基于时间维度局部特征的人体行为识别

摘要

视频人体行为识别算法中,局部特征三维模板卷积法难以避免背景中伪兴趣点且计算耗时.提出一种高效准确的融合时间维度和FAST角点特征的运动人体兴趣点检测方法,针对FAST角点不能表达时间维度信息的缺陷,将相邻三帧两两做差,然后在得到的前向和后向运动图像上进行FAST角点检测,取两个特征点集的交集作为当前帧运动人体局部兴趣点.该方法有效结合了时间维度信息和FAST算子的优点,具有耗时短、准确率高、运动相关性好的特点.最后应用词袋模型进行人体行为特征建模,分别应用SVM、KNN、决策树和LDA进行分类识别,在Weizmann、KTH数据库上进行测试,实验表明:SVM获得最好的分类性能,KNN获得最高的效率,因此KNN可以利用到实时的行为识别中.%In the human action recognition application,the traditional 3D template convolution method is time-consuming and difficult to avoid defects of pseudo interest points of background.To overcome this weak point,we propose a motion human local interest point detect method combining motion information and FAST feature.First,the difference is computed on every two adjacent frames,and the FAST(Features From Accelerated Segment) feature point detection is deployed on the two pieces of motion information by taking intersection of the two points set as final output with non-maximum suppression.With the low time-consuming FAST algorithm applied,this method should be an efficient motion intersection point detector with high accuracy and motion correlation.Finally,we use the BOW model to generate action feature vector.The classifier used is SVM (Support Vector Machine),KNN,Decision Tree and LDA.Performance is tested on deferent datasets,the simple KTH and Weizmann,SVM classifier,obtaining the best accuracy with KNN being more efficient.