基于贝叶斯数据融合的车载多类型障碍物探测和辨识方法研究

Next generation of Advanced Driver Assistance Systems (ADAS) is moving towards urban traffic scenarios. Compared to highway, urban traffic is much more complex because of cluttered road scenes and diversity of obstacles. The challenge of ADAS for use in urban traffic is developing advanced in-vehicle sensing systems able to accurately detect various obstacles from the cluttered scene and reliably classify them. It has been recognised that object classification is essential for safety applications in urban traffic to deploy appropriate protection systems or to make appropriate level of intervention. Existing researches on obstacle detection and classification design algorithms for a particular type of obstacles such as pedestrians or vehicles,therefore, are not able to detect and classify multi-class obstacles. Fundamentally differing from other researches which only target on single type of obstacles,this project aims to research on the method for detecting and classifying multi-class obstacles simultaneously, including pedestrian, vehicle, cyclist(motorcyclist) and traffic infrastructure. The basic idea is to use sensor fusion for obstacle detection and Bayesian network for obstacle classification since sensor fusion can provide multiple characteristics of obstacles and Bayesian network is suitable for modelling and inferring of an uncertain system. The key scientific questions to be answered by the project are: design of a Bayesian classifier for identifying multi-classes of obstacles, feature extraction based on sensor fusion, fusion of stereovision and optic flow, and positional calibration of multiple sensors. The project will lay a technical fundation for next generation of ADAS. The research outcomes will make significant theoretical contributions to the deployment of ADAS products with proprietary intellectual property rights.

下一代辅助驾驶系统将主要运用于路况复杂，多种障碍物并存的城区交通。用于城区的辅助驾驶系统应该不仅能够探测到多种障碍物，而且能够对其进行辨识，因为对于不同类型的障碍物，车辆应该选用的响应或保护措施是不同的。现有的对障碍物探测及辨识方法的研究都是针对某一种特定的目标(例如行人或车辆)来设计算法，不能同时探测和辨识多种类型的障碍物。本项目旨在引入可融合多种模式特征，适用于不确定性系统建模推理的贝叶斯网络技术，结合多传感器融合方法，研究面向城区环境的,具有通用性的障碍物探测及辨识方法。通用性是指能够同时探测并且辨识多种类型的障碍物，包括行人，车辆，骑车者以及交通设施。拟解决的关键科学问题包括：贝叶斯障碍物分类器，传感器信息融合的目标特征提取，立体视觉与光流的融合，三种传感器的空间标定。本项目的研究将为下一代辅助驾驶系统提供技术基础，对开发具有我国自主知识产权的辅助驾驶产品具有重要的理论意义。

对各种类型障碍物的探测与识别是智能汽车环境感知的重要内容。现有的车载障碍物探测及辨识方法的研究都是针对某一种特定的目标(例如行人或车辆)来设计算法，不能同时探测和辨识多种类型的障碍物。本项目研究了一种用于智能汽车的具有通用性的车载障碍物探测辨识方法。通用性是指可以对车辆、行人、路边交通设施等多类障碍物实现同时探测和辨识。在物体分割方面，提出和实现了一种采用立体视觉原理，基于物体位置的障碍物分割方法。在场景表达方面，提出和实现了借助U-视差特点，基于棒状像素的3D场景表达方法。在轮廓提取方面，对现有的蛇模型进行改进，提出并实现了一种基于距离向量卷积的活动轮廓模型，可以在噪声图像中准确提取被探测目标的几何轮廓。在类型辨识方面，提出并实现了基于动态贝叶斯网络模型的，综合多种形状特征的障碍物分类器。另外，本项目从理论上揭示了立体视觉和光流的内在关系，建立了光流、深度信息、相机自运动的数学模型，并以此为基础提出并实现了基于光流深度融合的车辆自运动参数估计方法，并重点解决了特征点中运动点的剔除问题。在完成相机自运动估计的基础上，进一步提出并实现了基于光流深度融合的运动目标探测方法。在多传感器融合方面，本项目研究并开发出了一种用于图像与激光雷达传感器位置标定的方法，可以将激光雷达探测到的点非常准确地对应到图像上的点。这些研究成果为下一代的智能车系统提供技术基础，对开发具有我国自主知识产权的辅助驾驶产品及智能车辆具有重要理论意义。