基于多因素贝叶斯的遮挡目标检测

摘要

Occlusion in complex scenes can result in loss of target appearance information, which leads to the loss of target in the detection process.By analyzing the dependence of the target representation on a particular layout, a multi-factor Bayesian method for occlusion target detection is proposed.First, a candidate proposal for the target local region is provided using the component model.Then, the visibility probability of the components is estimated according to the spatial layout, and the appearance characteristics and the deformation position of the target components are also considered.Finally, to complete the multi-factor weight learning,a Bayesian model based on appearance, deformation and visibility factors is constructed, and the area under the maximum curve is used to design the target detection and evaluation function.The experimental results demonstrate the effectiveness of the PASCAL data set, which outperforms the existing methods of target detection.%复杂场景中的遮挡现象会造成目标外观信息损失,致使检测过程中容易出现目标遗漏.通过分析目标表示对特定布局的依赖性,提出一种基于多因素贝叶斯的遮挡目标检测方法.首先,使用部件模型提供目标局部区域的候选提议,然后,根据空间布局关系估计部件的可见性概率,并同时考虑目标部件的外观特征和形变位置,最后,构建基于外观、形变、可见性因素的贝叶斯模型,并采用最大化曲线下方面积设计目标检测评价函数,完成多因素权重学习.实验结果证明在PASCAL数据集中的有效性,优于目标检测的现有方法.