摘要： 为实现基于计算机视觉的自动驾驶和高级辅助驾驶,需要对车辆前方的可行驶区域进行实时检测。可行驶区域的检测是图像分割问题,而目前主流的基于深度学习模型的Scene Parse方案,在实际应用中不能满足实时性要求。为此,构建一个超像素Appearance与3D特征融合的检测框架,实现从粗分割到细分割的两步分割流程。其中粗分割是基于RANSAC的快速平面估计,细分割则是基于粗分割路面估计概率的条件随机场模型,采用超像素进行加速。实验结果表明,该框架在Cityscape数据集上精确度和召回率超过90%,性能与SegNet,FCN 16 Scene Parsing相当,可满足X86平台和ARM平台的实时应用要求。

关键词: 自动驾驶, 高级辅助驾驶系统, 超像素, 双目视觉, 特征融合, 可行驶区域

Abstract: For automatic driving and advanced driving assistance task based on computer vision,the real-time detection of travelable area in front of the vehicle is necessary.The travelable area detection is an image segmentation problem,and the mainstream Scene Parsing scheme is based on deep learning model.In general,the deep learning framework is not real-time and cannot be applied in automatic driving task or advanced driving assistance task until now.Aiming at this problem,in this paper,an Appearance and 3D feature fusion detection framework is proposed,which implements two-stage process from coarse-grained segmentation to fine-grained segmentation.The coarse-grained segmentation is the fast estimation based on RANSAC fast plane,and the fine-grained segmentation is the conditional random field model based on the road estimation probability from coarse-grained segmentation,which uses super pixels to accelerate.Experimental result in the Cityscape data sets shows that the proposed framework achieves more than 90% accuracy and recall rate,which is comparable to SegNet and FCN 16 Scene Parsing framework.Its real-time applications can be achieved both in X86 and ARM platform.

Key words: automatic driving, Advanced Driver Assistance System(ADAS), superpixel, binocular vision, feature fusion, travelable area

中图分类号: 

• TP18