Road detection is fundamental for many applications, especially vision-based autonomous driving systems. To improve the accuracy of the detection results, most of previous research focus on designing feature encoders and classifiers. In this paper, a road detection method is proposed based on ensemble learning and key samples focusing. A road detection network is designed, which integrates classification results based on different feature combinations by weighted voting. The outputs of the network are further processed by morphological transformation. To focus on key samples, a novel loss function is proposed. The loss function can attach importance to hard samples and pay different attention to missed detection and false detection. The method is evaluated on KITTI dataset, and its effectiveness is verified.