基于单目视觉与深度神经网络的行为识别研究

Study on Human Behavior Recognition Based on Monocular Vision and Deep Neural Network

为解决多人场景下单目视觉行为检测难度大的问题,通过将单目视觉提取的多种特征集融合提出了一种新的行为检测模型。该模型通过深度卷积神经网络(Convolutional Neural Network,CNN)提取区域候选集,经过池化层获得单目视觉的感兴趣目标集;搜索决定人体姿势的25个骨架关键点,提取各个点的位置坐标向量。将2种特征融合传入CNN预测单目视觉的行为标签。实验结果表明,所提模型获得了较高的单目视觉行为检测精度,对多人场景的检测性能也优于其他对比模型。

To resolve the problem of behavior detection difficulty of monocular vision in the multi-person scene,multiple feature sets extracted from monocular vision images are fused to propose a new behavior detection model.Firstly,the region candidates are extracted through deep Convolutional Neural Network(CNN),and the regions of interest of monocular vision images are generated through the pooling layer;then,twenty-five skeleton critical points are located,the body pose is determined by the twenty-five points,and the position coordinate vectors of each point are extracted.Finally,both feature sets are fused,and then the results are transmitted to CNN to predict the behavior labels of monocular vision.The experimental results show that the proposed model achieves a high behavior detection precision of monocular vision still images;the detection performance is better than the other models in multi-person scenes.