卷积神经网络双目视觉路面障碍物检测

Stereo vision based obstacle detection using convolution neural network

针对双目视觉障碍物检测技术中,双目匹配精度差、障碍物检测算法稳定性与鲁棒性差等问题,对利用卷积神经网络的双目视觉障碍物检测方法进行研究。设计孪生卷积神经网络计算立体图像对的视差图;提出道路直线自适应阈值提取算法,精确提取V视差图中道路直线;利用光栅扫描法逐点判断像素点是否为障碍点,生成障碍检测图。实验结果表明,该方法能够生成精确视差图,有效提取V视差图中道路直线并实现路面障碍物精确提取,提高检测召回率和精确率。

In terms of stereo vision based obstacle detection technology,there are problems such as poor accuracy of stereo matching,poor stabilization and robustness,to solve these problems,an approach of stereo vision based obstacle detection based on convolution neural network was studied.Convolution neural network was applied to generate disparity of stereo image pair.A road line adaptive threshold extraction method was proposed to extract accurate road line in V-disparity map.The raster method was used to determine whether the pixels were obstacles or not,and obstacle detection map was generated.Experimental results show that the proposed method is effective in generation of precise disparity map,extraction of road line in V-disparity map for obstacle detection,improving the recall and precision rate.