Large Scale Road Scene Dense Semantic Mapping

Large scale road scene understanding is helpful for many applications, such as autonomous driving and unknown environment exploring. Nowadays, simultaneous localization and mapping is becoming mature, however, there are many problems about large scale semantic scene mapping, such as how to build the complete system properly, how to select good representation of 3D semantic map and how to calculate efficiently. Researchers pay much attention to the development of this technology. This thesis proposes a new solution and makes quantitative and qualitative experiments.This thesis proposes a complete large scale road scene dense mapping system. Firstly, it upsamples the 3D cloud with fused camera and lidar data, then, it takes advantage of stereo visual odometry to get the motion of camera, lastly, it splices together all dense frames to build the map.Because the 3D map reconstructed above cannot meet the requirements of related applications, this thesis realizes a global method under conditional random field which labels the whole map after building the 3D map. This method is workable and easy to understand, however, when getting a new frame, we need to rebuild the whole semantic map.To reduce the amount of memory and raise the efficiency of calculation, this thesis proposes a method which incrementally calculating semantic labels of the map with a conditional random field model. This method uses the means of 3D mapping proposed above and builds the semantic map by four stages as follows: firstly, detects newly built voxels, secondly, over-segments the points within these voxels into supervoxels, thirdly, labels these suervoxels under the guidance of neighboring frames, lastly, uses the rigid transformation matrix to fuse the newly labeled points with the already built map.Quantitative and qualitative experiments on KITTI dataset show that this approach can get an accurate large scale semantic map and decrease computational cost, at the same time, it can improve the labeling results at image level. Besides, the global method can get more stable results, however, incrementally calculating method is more efficient than the global method.