场景深度关系下的视频遮挡目标检测

Occlusion is an intrinsic phenomenon frequently appearing in realistic visual system. For this reason, the occluded object detection becomes more important and valuable in both theoretical and practical aspects with the application of intelligent video surveillance..The major obstacle is uncertainty estimation for complex occlusion relationships. Starting from object depth ordering associations in term of scene layout, we employ multiple visual cues to theorize occlusion inherence, and further use some constrains from physical projection to verify our generated occlusion. We construct novel object detection model based on occlusion relationship analysis..The detailed contents are listed as:.(1) By exploring the roles of appearance and motion cues for foreground object segmentation, we estimate moving object location in video..(2) By learning the depth ordering constrains form scene layout, we propose occlusion estimation method based on scene depth relationship..(3) By revealing the physical reason for object surface visibility, we put forward occlusion validation method according to visibility reasoning..(4) Finally, focusing on the mutual supports between segmentation object and detected in visual cognitive process, we explain the importance for three factors in evaluation of appearance, shape and coupled occluded pairs to measuring occluded object..Overall, we construct unified model towards occluded object segmentation and detection to boost the effectiveness and efficiency in video surveillance.

遮挡是一种普遍存在的视觉现象，遮挡目标检测对于智能视频监控具有重要的研究意义和应用价值。项目针对目标间复杂遮挡关系估计的研究难点，以场景布局下的目标深度排序关系为突破口，利用视觉感知线索揭示遮挡关系产生的内在成因，并利用物理投影约束条件描述目标间遮挡关系的合理性，建立基于遮挡关系分析的目标检测模型。具体研究内容如下：通过揭示外观和运动线索在前景目标分割的作用，估计视频中运动目标位置；通过研究场景布局对的目标深度排序的约束，提出基于场景深度的遮挡关系估计方法；通过揭示目标表面可见性的物理成因，建立基于可见性推理的遮挡关系验证方法；最终，通过研究视觉认知过程中目标分割与检测之间的相互支持关系，揭示外观评价、形状评价、遮挡关系评价对遮挡目标检测的重要性，构建面向遮挡目标分割和检测的统一模型，实现视频监控中遮挡目标的有效检测。

遮挡是一种普遍存在的视觉现象，遮挡也是目标检测、跟踪等领域的主要挑战。项目在分析遮挡关系产生的内在原因基础上，以遮挡检测和遮挡关系验证两个关键科学问题为核心，构建面向遮挡目标分割和检测的统一模型。.具体包括四个方面的内容：（1）在目标位置估计方面，提出了运动外观多线索的目标轮廓检测模型，在CMU数据集上达到并超过了近期研究结果；提出一种二进制光流描述子，提高了在Caltech目标提议的可靠性，并设计光照恢复模型，提高多线索特征提取的可靠性。（2）在遮挡检测方面，课题组提出的基于因子图场景深度线索表示与推理模型，在Make3D、BSDS数据集都达到并超过了近期研究结果。（3）在遮挡关系验证方面，研究了基于深度生成有向网络的区域间遮挡关系估计模型，在深度网络框架后端构建隐Markov场推理模型，验证遮挡关系，在Make3D、NYU数据集都达到并超过了近期研究结果。（4）在分析遮挡目标应用方面，研究了基于遮挡关系的目标分割与检测统一模型，并验证遮挡关系在目标跟踪、异常检测中的性能。.项目的主要科学意义在于遮挡成因的分析和遮挡线索的提取，体现了计算机视觉、生物视觉、物理学得学科交叉应用。项目组的研究成果，已经被成功应用于智能视频监控任务，实现企业成果转化。