基于车载鱼眼相机的运动目标检测关键技术研究

In recent years, fish-eye camera has been used gradually for on-board moving object detection due to its advantages of large imaging angle. However, it also brings imaging distortion problem which increases the difficulties of moving object detection. This project will study the imaging characteristics of fish-eye camera and moving object detection problems for vehicle environment. Firstly, the imaging law of fish-eye camera is analyzed and an anti-distortion optical flow extracting method is investigated to make the feature points’ detection and matching stable. Then, a deformation measure method is explored to distinguish flat surface from three-dimensional objects, and multi-feature constraint equation of motion parameters is set up in order to estimate the motion parameters under bumpy conditions. Based on this, the reverse projection error model of feature points is deduced to detect moving feature points in strong parallax scene. Finally, based on the theory of conditional random field, the potential functions of pixel points’ time and space consistency restriction are constructed to achieve moving objects’ accurate segmentation. The project involves some difficult problems in the field of moving object detection by monocular vision. The researching results of this project can be applied in the field of intelligent vehicle, driver assistance, robot navigation, etc. which is of theoretical significance and application value.

近年来，鱼眼相机由于具有成像视角大的优点，逐渐被应用到车载运动目标检测领域。然而，鱼眼相机也带来了成像的扭曲失真问题，增加了运动目标检测的难度。本项目围绕鱼眼相机的成像特点，对车载环境下的运动目标检测问题展开研究。首先，分析鱼眼相机的成像规律，研究抗扭曲的特征点光流获取方法，解决鱼眼图像的特征点稳定检测及匹配问题；然后，探究基于形变衡量的平面物立体物区分方法，构建运动参数的多特征约束方程，解决颠簸路况下的运动参数估计问题；在此基础上，推导特征点反向投影误差模型，解决车载强视差场景下的运动点检测问题；最后，基于条件随机场理论，构建点的时空一致性约束势函数，实现对运动目标的精确分割。本项目研究内容涉及单目视觉运动目标检测领域的难点问题，其研究成果可用于智能车，汽车辅助驾驶，机器人导航等，具有重要的理论意义和实用价值。

运动目标检测是实现智能交通，智能车，智能机器人的基础技术。本项目聚焦于解决基于车载鱼眼相机的运动目标检测关键技术问题。提出了基于球面归一化坐标的形变处理方法，并基于球面归一化坐标，推导了鱼眼相机运动参数估计及运动目标检测的相关理论。主要成果有：1）提出了基于形变衡量的地面和障碍物区分方法。提出了一种形变衡量方法，通过衡量图像区域的形变量，区分地面和障碍物。该方法无需使用物体外观特征，也无需使用运动参数信息，提高了地面和障碍物区分的准确性和鲁棒性。2）提出了基于多特征约束的运动参数估计方法。建立了地面特征和背景特征的统一约束方程来估计运动参数，解决运动参数相互干扰问题。通过使用多特征约束，提高了运动参数估计的稳定性。通过使用全参数模型，解决汽车在颠簸路况下的运动参数估计误差大的问题，提高了运动参数估计的准确性。3）提出了基于反向投影误差的运动特征点检测方法。基于球面归一化坐标，推导了适用于鱼眼图像的反向投影误差模型，通过融合高度约束及深度约束，减轻了单目视觉中的运动点检测退化问题。4）提出了基于多信息融合的运动目标分割方法。针对车载环境下，现有方法难以对运动目标进行完整分割的问题，提出了一种基于多信息融合的运动目标分割方法。将获得的运动特征点位置信息、深度图估计信息、地面及障碍物语义分割结果信息进行空间融合，将历史检测结果和当前帧检测结果，通过深度学习获得的稠密光流进行时序融合，最终实现了鱼眼相机上的运动目标分割。.本项目围绕基于鱼眼相机的运动目标检测关键技术问题进行研究，完成了预期目标和任务。发表论文5篇，其中4篇已被EI检索，本项目均为第一标注。申请相关算法专利6项，其中1项已经授权，并实现专利成果转化，另外5项，已经公开。培养硕士研究生4名。