针对移动平台有限的计算资源以及基于彩色图像的道路检测方法在极端光照情况下及路面类型变化时存在的不足,提出了一种融合彩色图像和视差图像的基于9层卷积神经网络的快速道路检测算法。提出一种数据输入层预处理方法,将视差图变换为视差梯度图以强化地面特征,降低网络深度需求。所提两种网络结构为双通道后融合网络和单通道前融合网络,分别用于卷积特征分析和快速道路检测。实验使用KITTI道路检测数据集并人为划分为普通和困难两组数据集,对该算法进行实验对比和分析,结果表明:与基于彩色图像的卷积神经网络方法相比,该算法在普通数据集上最大F₁指标(MaxF₁)提升了1.61%,在困难数据集上MaxF₁提升了11.58%,算法检测速度可达26 frame/s,可有效克服光照、阴影、路面类型变化等影响。

The road detection method based on color image exists problems under the extreme lighting conditions and changing road surface types, and the computing resource in moving platform is limited. So, based on the 9-layer convolutional neural network, a fast road detection algorithm is proposed to mix the color image and the disparity images. A new preprocessing method is applied in the data input layer, which can transform the disparity images to disparity gradient maps so as to enhance the representation of roads and reduce the demand for network depth. Two proposed networks are proposed including a double-path convolutional neural network which is used to analyze the characteristics of the convolutional neural network, and a single-path convolutional neural network which is applied to detect the road rapidly. The performance of the proposed algorithm is experimentally compared and analyzed on the KITTI road detection dataset which is divided into a common database and a difficult database artificially. The result demonstrates that, compared with the convolutional neural network method based on color images, the MaxF1 measures on the common database and difficult database improve by 1.61% and 11.58%, respectively, and the detection speed can be 26 frame/s. The proposed algorithm can overcome the impact of the lighting, shadow and the changing road surface effectively.