Study On Vehicle Environment Modeling Method Based On Stereo Vision

Booming development of Artificial Intelligence have a huge influence on vehicle industry during this period,the vehicle is becoming more and more intelligent,clean and safe.And self-driving vehicle is meaningful for improving the transportation safety,social benefits and economic benefits.In the paper,we explore and calculate the geometrical information,semantic information and pose information for transportation environment based on stereo vision sensor and we fuse these information together to build semantic maps to establish foundation for self-driving car behavior decision.At first,aiming to environment semantic information we propose a novel and compact fully convolutional neural network to obtain the pixel-wise semantic information.Simultaneously,we fuse the depth map and RGB image to obtain the 4-channel RGB-D images,we take RGB-D image as input of network to improve the accuracy of segmentation.The experiment results show the depth map is helpful for improve segmentation accuracy,and our network is competitive in speed and accuracy.The semantic information will be used in the chapter three to build semantic map.Secondly,we propose a novel semantic SLAM system to improve the ORB-SLAM2.Original ORB-SLAM2's map has no semantic information and just have sparse geometry information.We present a fourth dense point cloud map thread,accurate localization and simultaneously build geometrical model for vehicle by fusing geometry and semantic information to help vehicle distinguish the driving areas and obstacles.We also test the effectiveness and university of the algorithm by using the KITTI dataset and vehicle test.Finally,we calculate the pose of moving objects such as vehicle and people.In the chapter we present a novel vehicle detection way by fuse Tiny-Yolo and optical flow,and then calculate the pose of detection object by using ICP on the foundation of optical flow.We also prove vehicle detection way that we present is helpful for avoid leak detection and duplicate detection compared to Tiny-Yolo.