本项目旨在围绕复杂环境中视觉注意选择中信息反馈的科学问题，借鉴生物视觉系统的认知机理，探索视觉信息与先验行为知识协同的有效计算途径，将生物选择性注意、视觉反馈机制与任务分层知识协作引人到无人车辆交通场景感知与表达的计算框架中，研究大场景中显著性目标感知的关键技术，降低实时视觉信息处理的时间复杂度，提高机器感知的效率。构建多相机协同工作的车载视觉感知平台。建立完善的驾驶任务分层模型与多视点协同知识库。研究基于视觉注意的显著性目标检测模型。在利用目标的显著性、时空约束、以及目标与环境的动态差异性，结合概率图模型，研究鲁棒的显著性目标跟踪模型。利用统计学习，结合多类传感信息，建立多视点显著性融合的场景理解与表达模型。通过本项目的组织与实施可以为无人车在复杂的交通大场景中提供实时高效的显著目标感知，为驾驶决策提供及时的信息反馈，对一般环境中的场景感知研究具有重要的借鉴意义。

This proposal is aiming to study the key techniques of salient object perception in a wide scene, which are expected to reduce the computational cost in processing visual information and improve the efficiency of environment perception. The biologic selective attention, visual feedback scheme and task hierarchical knowledge will be introduced into the framework of traffic environment perception and expression for autonomous vehicle. Surrounding the problem of information feedback in the visual selective attention in complex environment, and referring to the cognitive mechanism in the biological vision system, we will explore the efficient computational algorithms in the cooperation of visual information and priori behavior knowledge. In this project, we will build an on-board visual perception platform which has multiple cameras working cooperatively, construct a hierarchical driving task model and multi-viewpoint cooperation knowledge base, and research on the salient object detection model based on visual attention. Making use of object saliency, space-time constraints and the dynamic difference between the objects and environment, we will also study robust algorithms for tracking salient object based on the probabilistic graphical model. We will construct multi-viewpoint saliency fusion based scene perception and expression model, using statistical learning techniques. The implementation of this project will provide real-time, high efficient salient object perception algorithms in a wide scene, which will give feedback to the driving decision module. Besides that, research on general scene perception can also benefit from it.