基于人体图像生成的姿态无关人物识别

Pose-Independent Person Identification Based on Human Body Image Generation

人物识别技术能够使机器人具备对用户身份识别的能力,从而有效提高机器人的智能交互水平。人物识别面临的主要挑战之一是姿态的变化对人物身份特征提取的影响。针对该问题,提出基于人体图像生成的姿态无关人物识别方法,通过生成与库中目标人物相同姿态的人体图像,消除姿态变化对人物外观特征造成的影响。该方法首先利用人体分割图将人体区域与背景分离,尽量降低复杂多变的背景对人物外观特征的干扰;然后在目标姿态的引导下生成与目标图像姿态一致的人物图像;最后设计了一个特征融合模块将源图像和生成图像的身份特征进行融合,提取姿态无关的鲁棒身份特征用于人物识别。此外,为更好地区分不同的人物,在训练中生成相同姿态的负样本,对约束模型学习更为细粒的可鉴别性身份特征。人物识别和人体图像生成的实验结果验证了该方法的有效性。

Person identification technology enables the robots to have the ability to recognize the identities of users,which effectively improves the intelligent interaction level of robots.One of the main challenges of person identification is the influence of the pose changes on person feature extraction.In order to solve this problem,a pose-independent person identification method based on humon body image generation is proposed,which aims to eliminate the influence of pose change on the person appearance features by generating the human body images with the same poses as the target persons in the dataset.Firstly,the method uses the human body segmentation map to separate the human body regions from the background to minimize the interference of the complex and changeable background on the human body appearance features.Then,a human body image with the same pose as the target image is generated under the guidance of the target pose.Finally,a feature fusion module is designed to fuse the identity features of the source and generated image to extract pose-independent robust identity features for person identification.In addition,to better distinguish different persons,negative samples with the same pose are generated in the training process to constrain the model to learn more fine-grained discriminative identity features.Experimental results on person identification and human body image generation demonstrate the effectiveness of the method.