任务相关视觉注意机制与非安全驾驶状态分析方法研究

Human driver inattention is the leading cause of traffic accidents . Active safety systems, which could be employed to alleviate the mentioned problem above, is currently on the threshold of a prototype system to the process of industrialization. And how to assess the mental state of the driver of such a system is difficult but decisive to make that the Active safety a practical and promising technology. Studies have shown that, in normal driving, human driver has the ability to quickly allocate their attention to safety related traffic events, and the distribution model has certain rules to follow. Thus, a criterion for driving safety status evaluation is that whether the driver pay attention to the related traffic events timely, and do not be attracted by non-related traffic events for a long while. Traditionally, unsafe driving status is mainly analyzed through fatigue detection using indirect, external fatigue features. However, fatigue is a gradual process, and when fatigue occurs, there must has been quite a long time that driving in the unsafe driving condition; Also, no fatigue occurrence does not guarantee current driving status is of safety. According to the former research foundation of human visual attention mechanism, and based on the study of intelligence driving in the recent years, this project plans to research on the modeling of driving-task-driven visual attention mechanism and fixation data analysis based attention allocation pattern and its inherent laws, to propose an direct, rapid and reliable method for unsafe driving status detection based on the fusion of visual attention computation and driver's fixation data analysis. These studies could probably be meaningful in the realization of future advanced driving assistant system and autonomous driving system, and provide new theories and key technologies for such systems.

人类驾驶员注意力不集中是导致交通事故的主因。目前用以缓解这一问题的主动安全系统正由原型系统阶段向产业化方向迈进。而如何评估驾驶员精神状态是制约此类系统走向实用的核心和难点问题。研究表明，人类在正常驾驶时可迅速将注意分配到与行驶安全相关的交通事件上，且分配模式有一定规律可循。因此，安全驾驶状态的判断标准是驾驶员是否及时注意到当前环境中相关交通事件、且没有长期被非相关事件吸引。传统对非安全驾驶状态分析主要是通过间接、外在的特征进行疲劳判断。然而疲劳是渐进过程，当检测到疲劳时驾驶员已有相当长时间处于非安全驾驶状态；另外，不疲劳并不意味着一定处于安全驾驶状态。本项目立足于视觉注意机制建模和智能驾驶领域的工作基础，探索驾驶任务驱动的注意模型，研究基于实测视线分析的驾驶员注意分配模式及其内在规律，从而建立注意计算与眼动实测相匹配的非安全驾驶状态分析方法，为辅助驾驶和自主驾驶技术储备理论基础和关键技术。

本项目针对车辆驾驶过程中，驾驶员驾驶状态以及车辆周边环境感知问题进行研究，研究了任务相关视觉注意机制与非安全驾驶状态分析方法。主要研究内容和成果包括：（1）研究了特征分层与注意机制建模，深入分析与研究了特征分层基于注意机制建模的方法，进一步探索了这些因素与最终注意形成之间的关系，研究了对驾驶员注意具有重要影响的特定物体（车辆、行人）检测方法，提出了级联神经网络和Grid Expending上采样方法；（2）进行了眼动数据的采集与分析，弥补了当前国际流行的显著性检测数据库在眼动数据和显著性区域相关性的缺失，使用改进的Intelligent Scissors算法对眼动数据进行了像素级别的高精度显著性区域标注，为进一步研究视觉注意机制提供了丰富数据支撑；（3）探索了结合眼动数据和实车道路图像显著性区域来判断驾驶员非安全驾驶状态的基本方法。其中研究了驾驶员预测注意与实测眼动数据的匹配方法，提出了一种基于全局及局部信息的显著性检测（GLRCNN）的端到端的深度网络模型，提高了抑制非显著性干扰的能力，并提出了一种驾驶员非安全驾驶状态检测系统方案，利用一种基于神经元活性描述方法的Neural Accumulation算法来触发告警；（4）提出了研究方案的系统实现方法，项目使用相机捕捉外界环境，同时使用眼动仪获取驾驶员的注意力状态。为了提高基于深度学习的注意模型的实时性，提出了一种扩展Haar滤波器的深度学习加速和轻量化算法，有效压缩了深度模型的计算量和存储量，为本项目提出的驾驶员非安全状态监测的部署奠定了基础。本项目的研究成果既包含任务相关视觉注意机制与非安全驾驶状态分析模型的理论创新又强调了算法的实际应用价值，相关成果成功地应用于无人驾驶平台，并表现出良好的性能，在中国智能车未来挑战赛、“跨越险阻”等比赛中取得了优异的成绩。本项目的研究对解决高级辅助驾驶系统的环境感知问题，以及推动我国无人驾驶技术的发展具有重要意义。