论文部分内容阅读
由于人体动作的多样性、场景嘈杂、摄像机运动视角多变等特性,导致人体动作识别的难度增加。为此,提出一种基于3D卷积神经网络结构的人体动作识别方法。以连续的16帧视频为一组,采用视频图像的灰度、x方向梯度、y方向梯度、x方向光流、y方向光流做多通道处理,有效地训练网络参数,经过5层3D卷积、5层3D池化增加提取特征中时间维度的动作信息,最终通过两层全连接与softmax分类器得到识别分类结果。通过与iDT、P-CNN、LRCN三种典型算法比较,实验结果表明,本文提出的方法识别准确率更高,且运行速度更快。
Due to the diversity of human actions, scenes noisy, changing camera perspective and other characteristics of the movement, resulting in increased difficulty identifying the human body movements. For this reason, a human body movement recognition method based on 3D convolution neural network structure is proposed. Taking 16 continuous video frames as a group, multi-channel processing is done by using video image gradation, x-direction gradient, y-direction gradient, x-direction optical flow and y-direction optical flow to effectively train the network parameters. Product, and 5-layer 3D pooling to extract the action information of the time dimension in the extracted features, finally, the classification result is identified through two full connections and the softmax classifier. Compared with the three typical algorithms iDT, P-CNN and LRCN, the experimental results show that the proposed method has a higher recognition accuracy and faster running speed.