辅助驾驶车载视频信息的结构场模型与理论研究

项目名称： 辅助驾驶车载视频信息的结构场模型与理论研究

项目编号： No.61273237

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 谢昭

作者单位： 合肥工业大学

项目金额： 76万元

中文摘要： 在辅助驾驶的应用背景下，分析车载视频中关于分割、检测、深度和速度估计的计算机视觉问题及路况状态预测问题，将各个任务模块置于结构场模型的框架内，定义不同类型的场能量形式，充分考虑效率性、准确性、鲁棒性和灵活性，用统一的学习推理方法求解子问题。研究视频二值分割的时序二阶关系，在时空场上分析超像素图切割推理方法；构建目标检测时空场和语义标记场表达，在语义关联和空间约束下平滑连续参数的求解过程，用置信度传播方法推理类别标记；分析深度空间场的尺度关联特性，实现尺度平滑策略下参数和变量的梯度似然求解方法；讨论运动光流场下局部区域速度的一致性，引入正则化先验约束，用随机梯度方法近似场能量似然项进行速度分解估计；采用二值检测向量描述路况状态场，定义不同状态的跳转矩阵，根据计算出的目标位置、运动和深度信息得到当前路况的似然估计，通过置信度传播预测路况状态信息并给出控速决策。

中文关键词： 车载视频；结构场模型；学习推理；生物视觉；运动表达

英文摘要： With application of assistant driving, we will draw intensive attentions to several critical issues in on-board video parsing, which can be naturally decomposed into some matters in computer vision as segmentation, detection, depth-and-velocity estimation and prediction problem of road situation. Unifying each task into structural field models, we should make efforts to achieve well-defined energies with typically distinct representations and regularize the learning and inference methodologies with sufficient considerations of efficiency, accuracy, robustness and flexibility. Separately, in video binary segmentation, we will focus second-order sequence in spatio-temporal field for graphcut-based inference in super-pixels; coupling with semantic labeling field in category detection, we intend making optimizations by smoothing continuous-parameter and propagating labels' belief for prediction in contextual constraints; associating with scale-smoothing properties in depth field, we should solve the tractable gradient of likelihood in newly-defined energies; and then the consistency in moving flow field should be maintained by introducing regularization as prior constraints, we may use stochastic gradient method to estimate two orthogonal sub-velocities; finally, the road situation should be represented as binary-de

英文关键词： on-board video；Structure filed model；learning and inference；biological vision；motion representation