基于自动编码器与概率神经网络的人体运动行为识别方法
2018-02-01李兴肖秦琨
李兴+肖秦琨



摘要:
为了准确而快速地识别出图像中人体的某种行为特征,提出了一种全新的三层自动编码器与PNN概率神经网络结合的网络模型。把人体行为的关键帧从包含10种人体行为的视频库中提取出来,用背景减除法提取人体轮廓图并进行二值化图像处理,根据时间序列叠加轮廓图,组成含有10个特征轮廓的数据库,随后将特征轮廓数据在自动编码器中编码,编好码的特征数据随即进入概率神经网络进行训练学习。将训练好的神经网络对这10种人体行为进行准确识别。实验结果表明,经过自动编码器处理后的人体特征进行PNN神经网络识别,比单纯用BP神经网络识别准确率提高5%以上,由此证明该方法有效可行。
关键词:
人体行为识别;自动编码器;PNN概率神经网络;BP神经网络
DOIDOI:10.11907/rjdk.173068
中图分类号:TP301
文献标识码:A文章编号文章编号:16727800(2018)001001103
Abstract:In order to identify the behavior characteristics of the human body accurately and quickly, this paper proposes a new network model combining the three layer automatic encoder with the PNN probabilistic neural network. The first step of the key frames extracted from human behavior includes 10 kinds of human behavior video library, the second step with the background subtraction method to extract the body contour map and binarization image processing, according to the time sequence superimposed contour map containing 10 component feature contour database, with the data after feature contour in the auto encoder encoding, coding and feature data code immediately into the probabilistic neural network learning and training, the trained neural network is good for accurate identification of these 10 kinds of human behavior. The experimental results show that the recognition accuracy of PNN neural network is 5% higher than that of the pure BP neural network, and the validity and feasibility of the method are verified.
Key Words:human behavior recognition; autoencoder; BP neural network; probabilistic neural network
0引言
人體姿态识别在智能视频监控、动态人机交互、无人汽车驾驶等[12]领域应用广泛,是模式识别的一大热门研究方向。人体运动在时间和空间上都很复杂,因此需要有一个前期处理的清晰图像,目的在于提取有效的运动特征。
视频帧进行处理的方法有3种:背景剪除法、差分法和光流法,检测出人体行为区域,然后将该区域作为研究对象,从中获得人体行为特征[35]。Bobick和Davis 提出的基于视频的轮廓提取方法[6],剪除从背景场景中获取的有效轮廓特征,从视频中提取人体行为轮廓图。因为二值图像(Binary Image)按名字理解只有两个值:0和1,0代表黑,1代表白,或者说0表示背景, 1表示前景。……
