路面标识为无人驾驶提供重要的道路视觉信息，路面标识的正确识别是行车安全的前提。由于实际道路场景的复杂多变性，传统路面标识检测方法在嵌入式平台下算法鲁棒性和实时性方面仍面临着一些挑战。本文提出了嵌入式平台Jetson TX2下路面标识检测网络模型及优化算法：首先给出了基于层合并的区域全卷积网络R-FCN(Region-based Fully Convolutional Networks)的简化模型，实现了路面标识的高精度检测；接着，为了满足实际应用中实时推理需求，将简化的R-FCN网络模型部署在嵌入式平台NVIDIA Jetson TX2上，构建了基于TensorRT的模型推理优化加速方法，在嵌入式平台上实现了快速准确路面标识算法。提出的算法在自建路面标识库和相应的公开数据库进行了测试，实验结果验证了算法的有效性。

The road marking provides important road visual information for autonomous driving, so the correct identification is a prerequisite for driving safety. Due to the complex variability of real road scenes, the traditional road marking detection methods still face some challenges in terms of algorithm robustness and real-time performance of the embedded platform. So this paper proposes the road marking detection network model and optimization algorithm deployed on the embedded platform TX2: firstly, the high-precision detection of road marking is realized based on layer integration to simplify R-FCN(Region-based Fully Convolutional Networks); then, in order to meet the real-time reasoning requirements in practical applications, the simplified R-FCN network model is deployed on the embedded platform NVIDIA Jetson TX2, and the model inference optimization acceleration method based on TensorRT is proposed, which accelerates the inference phase of the network model. A fast and accurate road marking detection algorithm is implemented on the embedded platform. The proposed algorithm is verified in the self-built road marking database and the related public dataset, and the experimental results proved the effectiveness of this algorithm.