基于视觉一致性约束的监控目标精确检测研究

Accurate surveillance object detection is one of the key science problems that should be resolved urgently for the deep applying of intelligent video surveillance systems. In order to solve the problem of insufficient object detection accuracy, which is induced by the problem of visual-incoherence in the real complex surveillance systems, we try to develop two novel models: By integrating and extending of existing exemplar-based object detection model and category-based statistical detection methods, one novel object detection model based on visual consistency constraint will be designed firstly, which is used to transform the visual-incoherence object detection to the visual-coherence sub-category object detection, such as the shortcomings of insufficient detection accuracy in the typical statistical-based methods can be overcame. The second study is focused on the deeply-supervised network modeling, where domain knowledge is introduced to enhance the optimization of SVM model in each hidden layer of the typical convolution neural network. By this way, one new deeply-supervised learning model, which is drove by the data and domain knowledge jointly, can be developed to enhance the learning efficiency of deep CNN. Through the research on the visual consistency constraint based object detection modeling, deeply-supervised network learning drove by the data and domain knowledge jointly, the problem of visual-incoherence can be reduced effectively, and thus the object detection accuracy and robustness can be improved greatly. This project can be expected to play a positive role in improving the basic research level of intelligent video surveillance system.

监控目标的精确检测是目前智能视频监控系统深度应用面临的一个迫切需要解决的关键科学问题。本项目针对视频监控系统的复杂场景下目标视觉不一致对目标精确、鲁棒检测带来的挑战，一方面将基于范例的相似性检测模型与基于统计的检测模型相融合扩展，研究基于视觉一致性约束的目标检测方法与模型，将视觉不一致目标的检测转换为视觉一致的子类目标检测，以克服传统的统计检测模型存在的视觉不一致带来的检测精度不足等问题。另一方面，将深度卷积神经网络与领域知识辅助的监督学习相融合，提出一种能够支持知识与数据联合驱动的新的深监督学习模型,以领域知识的充分利用来提升深度模型的学习效能。通过上述两个方面的研究，力争有效解决监控目标检测中因视觉不一致造成的检测精度和鲁棒性不足问题，获得监控视频目标检测精度和鲁棒性上的性能突破，为提高我国智能监控视频的基础研究水平起到积极的促进作用。

监控目标的精确检测是目前智能视频监控系统深度应用面临的一个迫切需要解决的关键科学问题。本项目针对视频监控系统的复杂场景下目标视觉不一致对目标精确、鲁棒检测带来的挑战，研究提出了一系列基于视觉一致性约束的目标检测方法与模型，尤其在监控行人检测领域，提出了基于有价值上下文挖掘的行人检测模型，以及基于两级注意力的行人检测模型、基于多视角融合的行人检测模型，有效地提升了行人检测的性能；同时，通过将深度卷积神经网络与领域知识辅助的监督学习相融合，提出了多个能够支持知识与数据联合驱动的新的目标检测识别计数等模型：针对密集场景下目标检测计数问题，提出了一种密度图局部一致性约束的密集目标计数模型，以及一种多任务辅助的密集目标计数模型，有效提升了密集目标计数的性能；针对图像高级理解中的目标视觉关系检测识别问题，提出了视觉与文本联合驱动的目标视觉关系检测模型，通过融合文本与视觉特征，有效提升了视觉目标关系检测的性能。通过上述研究，为监控目标检测中因视觉不一致造成的检测精度和鲁棒性不足问题，提供了有效的解决方案，获得监控视频目标检测精度和鲁棒性上的性能提升，为提高我国智能监控视频的基础研究水平起到了积极的促进作用。