基于深度网络特征提取与核非线性分类的视频行为识别

摘要

行为识别是视频分析的重要任务.本文利用稀疏自动编码机和卷积神经网络这两种典型的深度网络从视频运动历史图像中提取行为特征,利用最优泛化核非线性分类器对特征进行分类,实现行为识别.实验结果表明,与常用的特征提取和分类算法相比,所探讨算法正确识别率更高;而最优泛化核非线性分类器比常用的分类器能更好地兼顾识别效果和效率.%Human action recognition is an important task in video analysis.In this manuscript,we try to use two typical deep networks to extract abstract features automatically from the motion history images of video samples and then classify these features using a kernel-based nonlinear representor.Experimental results show that the proposed algorithm outperforms traditional ones using HOG feature and k-Nearest Neighbor classifier in accuracy,and has approximate accuracy of SVM classifier but is more efficient.