论文部分内容阅读
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26frame/s,可有效克服光照、阴影、路面类型变化等影响。
In view of the limited computing resources of mobile platform and the shortage of road detection method based on color image under extreme lighting conditions and the change of pavement type, a fast road based on 9-layer convolutional neural network is proposed, which combines color image and parallax image. Detection algorithm. A data input layer preprocessing method is proposed to transform the disparity map into a parallax gradient map to enhance the ground features and reduce the network depth requirements. The two proposed network structures are dual-channel post-fusion network and single-channel pre-fusion network, which are respectively used for convolution feature analysis and fast road detection. Experiments using KITTI road test data sets and artificially divided into two groups of ordinary and difficult data sets, the experimental comparison and analysis of the algorithm, the results show that: compared with the color image based convolutional neural network method, the algorithm in the ordinary data set MaxF1 increased by 1.61%, MaxF1 increased by 11.58% on hard data set, and algorithm detection speed can reach 26frame / s, which can effectively overcome the influence of light, shadow and pavement type change.