CAJ | 학술논문

为了推广神经网络在人体动作识别中的应用，设计了一种基于分层识别框架和增强算法的动作识别系统，该系统融合了光流直方图、有向梯度直方图、Hu的矩特征、分块剪影和自相似矩阵等多种特征。为了与反向传播网络的增强相匹配，将传统的二分类增强算法扩展到多分类版本。此外，系统采用了包含预判决和后判决的分层识别框架，前者通过分析运动显著区域的位置，把动作粗分为几个子类，后者则利用额外的特征进一步提高识别准确率。基于Weizmann和KTH数据库的实验结果表明：神经网络相对于常用的支持向量机具有明显的优越性；结合分层识别的反向传播增强算法可以极大减少运算代价与动作类间的混淆，识别准确率较高。
위료추엄신경망락재인체동작식별중적응용，설계료일충기우분층식별광가화증강산법적동작식별계통，해계통융합료광류직방도、유향제도직방도、Hu적구특정、분괴전영화자상사구진등다충특정。위료여반향전파망락적증강상필배，장전통적이분류증강산법확전도다분류판본。차외，계통채용료포함예판결화후판결적분층식별광가，전자통과분석운동현저구역적위치，파동작조분위궤개자류，후자칙이용액외적특정진일보제고식별준학솔。기우Weizmann화KTH수거고적실험결과표명：신경망락상대우상용적지지향량궤구유명현적우월성；결합분층식별적반향전파증강산법가이겁대감소운산대개여동작류간적혼효，식별준학솔교고。
To popularize the application of neural network in human action recognition,an action recognition system based on the hierarchical recognition framework and the boosting algorithm is de-signed,which mixes together multiple features such as histograms of optical flow,histograms of ori-ented gradients,Hu’s moments,block-silhouettes and self-similarity matrices.To fit with the boos-ting of back-propagation (BP)networks,the standard binary AdaBoost algorithm is extended to a multiclass version.Besides,this system adopts a hierarchical recognition framework consisting of pre-decision and post-decision.The former can roughly classify the actions into several subcategories by analyzing the locations of motion salient regions,whereas the latter exploits extra features to fur-ther enhance recognition accuracy.The experimental results on Weizmann and KTH datasets show that neural networks exhibit obvious advantages over the popular support vector machine.The BP-AdaBoost algorithm combined with hierarchical recognition can greatly reduce the computational cost and confusions among actions to achieve high recognition accuracy.