摘要
建立特征的帧间运动对应是图像序列分析的关键性任务之一。本文将运动对应问题分解为特征对应和轨线延伸两个子部分,并将它们表述为一系列的费用最小化问题,用二维Hopfield网络进行求解。其中对解的约束均被明确地包含在费用函数中。Hopfield网络为运动对应问题的求解提供了一个简单、灵活而鲁棒的快速实现框架。该算法可进一步改进以处理遮挡问题。
Establishing motion correspondence of features between frames is a crucial task of image sequence analysis. In this paper we divide the problem of motion correspondence into two subsections-feature correspondence and trajectory extension and formulate them as a series of cost minimization problems that are solved using two-dimensional Hopfield neural networks. All the constraints on the solution are explicitly included in the cost functions. Hopfield network provides a simple, flexible, robust, and fast-implemented framework for motion correspondence. The algorithm can also be modified to take care of the occlusion case.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
1998年第4期390-395,共6页
Pattern Recognition and Artificial Intelligence
关键词
计算机视觉
运动对应
神经网络
解
Image Sequence, Feature Correspondence, Trajectory Extension, Hopfield Network
