基于卷积神经网络的语义同时定位以及 地图构建方法

环境感知技术是智能车功能实现的前提，而在感知的基础上提高智能车对环境的认知能力是实现全自动驾驶的关键。本文针对室外交通场景，基于卷积神经网络提出了一种新的智能车同时定位以及语义地图构建方法，对智能车进行定位，并且构建稠密的3D语义地图，提高智能车的环境感知、认知能力。首先，基于双目ORB-SLAM提出了一种四线程的双目SLAM（Simultaneous Localization and Mapping）方法构建稠密的3D点云地图，四线程分别为追踪线程，局部地图构建线程，回环检测线程以及稠密地图构建线程；其次，为提高智能车的环境认知能力，使用端对端的方法对图像进行语义分割，并且为提高语义分割精确率，将环境的几何信息也作为卷积神经网络输入；最后，将感知的能力与认知的能力相结合构建语义地图，为智能车实现全自动驾驶奠定基础。本文将算法在KITTI数据集上进行测试，整体算法速度为10帧/秒，语义分割的全局精确率为73.1%，构建的语义地图显示本文提出算法能够在大规模室外场景下重构全局一致性地图，并且帮助智能车实现对环境的解析。

Semantic simultaneous localization and mapping method based on convolutional neural network

The technology of the environment perception is foundation for the autonomous vehicles, but to realize the fully self-driving it is very vital to improve the cognition ability of autonomous cars. In this paper, we present a novel semantic Simultaneous Localization and Mapping way based on convolutional neural networks (CNNs), locating the position of vehicles and building the dense 3D semantic point cloud map when vehicles are outdoors. Firstly, we present a 4 thread SLAM method to build dense sematic 3D map including tracking thread, local mapping thread, loop closing thread and dense mapping thread based on original ORB-SLAM; Then to improve the cognition ability we obtain the pixel-wise sematic information by using an end-to-end deep learning way; Finally, we fuse perception and cognition abilities to build dense semantic map. We tested our method on the KITTI dataset., the whole speed of system is about 10Hz and the global accuracy of the convolutional neural network is 73.1%. Results indicate the system is effective in the real-time building a globally consistent semantic map, and at same time it helps vehicle perceive the environment.