面向智能视频监控系统中目标理解的长时程深度学习模型研究

项目名称： 面向智能视频监控系统中目标理解的长时程深度学习模型研究

项目编号： No.61471206

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 孙宁

作者单位： 南京邮电大学

项目金额： 81万元

中文摘要： 基于视频的目标理解，是对视频中目标属性和目标行为在语义层次上的描述，是视频监控系统智能化应用的重要体现。由于采集视角，目标姿态，光照条件等方面差异，加上之前目标检测和目标跟踪处理带入的误差等因素的影响，基于短时程图像序列的分析算法很难在实际条件下进行准确和稳定的目标理解。针对上述问题，本项目以深度学习理论为基础，从长时程图像序列中目标的时空相关特性出发，将卷积神经网络(CNN)和深度置信网络(DBN)进行融合，建立长时程深度混合神经网络(LDHNN)模型，利用CNN实现对三维图像序列数据的特征化和向量化，堆叠DBN来加深网络的层次，提升网络对目标长时程时空特征的学习能力。基于逐网络逐层的思想，推导LDHNN的训练算法，并利用多GPU并行运算来大幅加速训练过程，实现基于LDHNN的目标理解功能，为研究面向长时程图像序列的深度学习模型，提升现有视频监控系统的智能化应用探索一条有效的技术途径。

中文关键词： 视频语义理解；深度学习；长时程；深度混合神经网络；目标识别

英文摘要： Object understanding based on video, which is the semantic description of target attribute and behavior, and is an important embodiment of intelligent video surveillance system application. As the impact of video resolution, viewing angle, target poses and illumination dynamic, as well as the error of previous target detection and target tracking, the short-term image sequences based algorithm is unable to understand the object precisely and stably. In this project, a Long-term Deep Hybrid Neural Networks (LDHNN) model will be built with Convolution Neural Networks (CNN) and the Deep Belief Networks (DBN) based on the theory of deep learning and the principal of target spatial-temporal correlation in long-term image sequences. In the model of LDHNN, 3D image sequences are mapped to 1D feature vectors by CNN, and the stacked DBN deepen the level of entire networks. The network-wise and layer-wise training procedure of LDHNN can be accelerated obviously by parallel computing using multiple GPUs. It will provide the theoretical and technological support for long-term object understanding in video surveillance system.

英文关键词： video understanding;deep learning;long-term;deep hybrid neural network;object recognition