关键词: 双目视觉, 立体匹配, 图像分割, 三维重建, 深度学习, GPU并行计算

Abstract: Binocular stereo vision technology has always been of great significance in the field of computer vision research.Unlike monocular or multicular technology,binocular stereo vision has the advantages of low cost,high versatility,simple use and so on while it can accurately obtain the image depth.The three-dimensional perception technology based on binocular vision can greatly improve the computer's understanding and interaction ability to the real world,further enhance the adaptability of computer vision technology in complex and changeable scenes,and play an important role in the fields of automatic driving,robot navigation,industrial inspection,aerospace,etc.This paper focuses on 3D reconstruction and object perception technology in dynamic scenes.In most cases,dynamic objects in the field of vision usually need to be focused on,while static objects,especially the background and static objects in the scene that occupy the main space in most cases,can be ignored,but they do occupy a lot of resources in the actual calculation,It is obviously meaningless and inefficient to spend too much computing resources on targets that are not concerned in the scene.In order to solve this problem,based on the in-depth study of the current mainstream binocular stereo matching methods,image segmentation and other methods,this paper proposes a dynamic scene 3D perception technology based on binocular estimation.The main innovations and research achievements include:Aiming at the low cost and efficiency of the traditional binocular stereo matching algorithm in pixel by pixel computing aggregation,a binocular stereo matching method based on two-dimensional scene instance segmentation is proposed,and the target image after mask segmentation is used for stereo matching,which not only improves the matching performance but also reduces the difficulty of dynamic target matching.At the same time,in order to solve the problem of insufficient segmentation accuracy,the mask edge filtering optimization method based on rgb image is introduced to improve the efficiency and the reconstruction accuracy of the field of view point cloud.Secondly,real-time target point cloud production is carried out based on binocular estimation depth learning network,and a real-time dynamic target perception algorithm based on GPU accelerated neighboring frame point cloud is proposed.At last,a two-dimensional and three-dimensional dynamic object real-time perception technology is proposed,which can quickly recognize the dynamic object in the detection environment while realizing real-time three-dimensional reconstruction of the target scene.

Key words: Binocular vision, Stereo matching, Image segmentation, 3D reconstruction, Depth learning, GPU parallel computing