基于CNN与VLAD的人体行为识别研究
首发时间:2018-03-09
摘要:人体行为识别是很多实际应用中的关键技术,例如视频监控、人机交互、虚拟现实和视频检索等,所以智能的人体行为识别技术有着很高的研究价值和应用前景,特别是如今视频数据呈爆炸式增长的时期。深度学习是现阶段十分火热的研究领域,尤其是CNN模型已经在图像识别上取得了巨大的成功,相较于传统的人工特征方法,CNN特征在对原始数据的特征表达方面有着绝对的优势,已经逐渐地取代了以HOG和SIFT等人工特征成为了主流的特征提取方法。VLAD是一种能将尺寸不同的视频特征数据表示成尺寸相同的特征向量,通过VLAD特征表示可以使得视频特征数据满足一般分类器的输入要求并得到分类结果。本文提出了一种CNN特征提取结合VLAD特征表示的人体行为识别方法,并在Youtube数据库上进行验证,最终取得较高的识别准确率,证明了CNN结合VLAD是一个十分有效的人体行为识别方法。
关键词: 人体行为识别;深度学习;CNN;VLAD 特征提取 特征表示
For information in English, please click here
Study on human behavior recognition based on VLAD and CNN
Abstract:Human behavior recognition is the key technology in many applications, such as video surveillance, human-computer interaction, virtual reality and video retrieval, so human behavior recognition technology intelligent has a very high research value and application prospect, especially in today\'s video data explosive growth period. Deep learning is a research field at this stage is very hot, especially the CNN model has achieved great success in image recognition, artificial features compared with traditional method, CNN feature has an absolute advantage in the expression characteristics of the original data, has been gradually replaced by HOG and SIFT as artificial feature the mainstream feature extraction method. VLAD is a video feature data of different sizes into the same size feature vector, through the VLAD feature representation can make the video data meet the general classifier requirements and classification. This paper presents human behavior recognition method based on VLAD feature extraction a CNN feature, and verified in the Youtube database, finally achieved a higher recognition accuracy, and prove that CNN combined with VLAD is a very effective method for human behavior recognition.
Keywords: human behavior recognition deep learning CNN VLAD feature extraction feature representation
基金:
引用
No.****
同行评议
共计0人参与
勘误表
基于CNN与VLAD的人体行为识别研究
评论
全部评论0/1000