基于立体视觉深度学习的车辆前方可通行性分析研究

项目名称： 基于立体视觉深度学习的车辆前方可通行性分析研究

项目编号： No.61203171

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 自动化学科

项目作者： 李琳辉

作者单位： 大连理工大学

项目金额： 25万元

中文摘要： 在城、郊区道路环境中，基于车载环境感知系统分析车辆行驶方向上的可通行性，是实现安全预警的关键。现有的方法以障碍检测为主，缺乏对环境深入的理解，且难以自动适应新的环境。为此，本项目立足立体视觉传感器，从模拟双眼感知的角度提出离线学习和在线学习相结合的可通行性分析方法。首先，按可通行程度对环境分类，确立视觉理解的目标；然后，模拟大脑及视觉皮层细胞的认知过程，探索基于深信度卷积神经网络的认知结构，适合多维图像直接输入，具有权值共享、可深度学习等特点。在离线学习部分，研究无监督逐层训练和有监督微调相结合的半监督深度学习方法，获取接近目标本质的特征描述，并减少标记样本的数量；在线学习部分，基于近景重建和车辆当前位置可通行性分析，提出由近及远的自监督环境分类方法，使车辆能够自动"熟悉"未知环境，实现可靠的安全预警。最终，构建通用性强、可扩展的车载环境感知一般求解框架，推动安全辅助驾驶系统的智能化发展。

中文关键词： 车载环境感知；可通行性分析；立体视觉；深度学习；深信度卷积网络

英文摘要： Vehicle's forward traversability analysis based on the onboard environment perception system is the key to achieve safety forewarning in urban or suburban environment. Most current methods just can distinguish obstacles from environment, which lack of deep understanding and can't adapt to new unknown environment automatically. Then, this project use the stereo vision sensors, and puts forward an off-line learning and on-line learning based traversability analysis method by simulating the eyes' perception process. Firstly, the targets of visual perception are confirmed based on the environment traversability classification. Secondly, the brain and visual cortex cells' perception process are simulated by establishing a Deep Belief Convolutional Nets (DBCN). The Net is fit for multidimensional images directly input and deep learning, also has the advantages of incorporate weights sharing. In the off-line learning part, semi-supervised learning method is proposed to describe the targents' basic feature and avoid large labeled samples. Each layer of DBCN are pre-trained independently and sequentially in unsupervised mode, and then all the parameters are fine-tuned in supervised mode. In the on-line learning part, based on the close view reconstruction and vehicle's current position's traversability analysis, we propo

英文关键词： Onboard environment perception；Traversability analysis；Stereo vision；Deep learning；Deep Belief Convolutional Nets