仿视觉搜索机制的多视认目标驾驶环境感知方法研究

This project is oriented to the precise and real-time environmental perception requirements for autonomous vehicles, aiming at solving the problem of low detection precision and high space-time complexity of multi-target environmental perception method, to explore the multi-target driving environmental perception method guided by the driver visual search mechanism. By analyzing the spatial and temporal distribution characteristics of drivers’ gaze points, and the changing characteristics of drivers’ eye movement in time series, a visual search region sub-model under the driving environment containing different types of single-target and a visual search strategy sub-model which can represent targets’ recognition order in multi-target driving environment will be built. The part model will be built on the basis of the visual search region sub-model, and the single-target detection method will be optimized combining machine learning architecture with the part model to improve the detection precision of obscured targets. Base on the visual search strategy sub-model, the multi-target detection hierarchy architecture will be constructed. A multi-target detection model, based on visual search mechanism and a multi-target driving environment perception method and system with higher visual accuracy and less space-time complexity, will be developed by embedding the single-target detection method as a functional node into the multi-target detection hierarchy architecture. This project can give full play to the spatial and temporal regularity advantages of driver visual search mechanism, meet the precision and speed requirements of multi-target driving environment perception, and improve the stability and safety of autonomous vehicles.

项目面向自动驾驶车辆环境感知的准确性和实时性需求，针对多视认目标感知方法检测精度较低和时空复杂度较高的问题，研究模仿驾驶人视觉搜索机制的多视认目标驾驶环境感知方法。项目通过分析驾驶人视线点时空分布特征和时序下驾驶人眼动参数的变化特征，构建单视认目标驾驶环境下的视觉搜索区域子模型，以及表征多目标驾驶环境下目标视认顺序的视觉搜索策略子模型；以视觉搜索区域子模型为指导构建部件模型库，结合机器学习架构优化单目标检测方法以提高被遮蔽目标检测精度；基于视觉搜索策略子模型构建多目标检测层级架构，将单目标检测方法作为功能节点嵌入多目标检测层级架构，建立仿视觉搜索机制的多目标检测模型，进而开发具有较高视认精度和较小时空复杂度的多视认目标驾驶环境感知方法和系统。项目可充分发挥视觉搜索机制空间上的针对性优势和时间上的规律性优势，满足多视认目标驾驶环境感知的精度需求和速度需求，提升自动驾驶车辆的稳定性与安全性。

项目面向自动驾驶车辆环境感知的准确性和实时性需求，针对常规多视认目标感知方法检测精度较低和时空复杂度较高的问题，研究模仿驾驶人视觉搜索机制的多视认目标驾驶环境感知方法。.项目完成了多变量耦合状态下的视觉搜索机制研究。基于心理学、认知科学、概率图模型等学科理论揭示了驾驶人视觉搜索机理，提出了基于双目视觉的车用障碍物检测方法，实现了对真伪障碍物的准确判别；提出了视觉搜索区域的确定方法、装置及驾驶模拟器，实现了驾驶环境感兴趣区域准确、快速的识别；提出了多目标检测网络的构建方法、多目标检测方法及装置，实现了高效、高精度的多目标检测。.项目完成了仿视觉搜索机制的感知模型构建方法研究。面向多变量耦合状态下的汽车自动制动系统环境感知需求，分析了单视认及多视认目标驾驶环境下的视觉搜索策略，提出了驾驶环境的识别方法、装置及模拟驾驶器，实现了面向环境感知的驾驶人视觉搜索机制的重构；构建了面向不同类型单视认目标及多目标驾驶环境感知的视觉搜索区域子模型，基于视觉搜索区域子模型和机器学习架构优化了单目标检测方法，提出了模仿驾驶人视觉搜索机制的驾驶环境感知方法体系。.项目完成了仿视觉搜索机制的线控系统研发。结合雷达和相机的AEB系统、识别道路限速标识读取限速信息控制加速的ISA系统、对比车辆当前位置以及目标轨迹得到的横向误差和航向误差进行修正转向控制的自动转向系统，进行制动、加速、转向功能的集成，完成自动驾驶巡逻车的研发。