Study On Depth Extraction And Colored 3-D Reconstruction Of Night Vision Scenes

In recent researches, computer vision technologies are always used in daylight or under artificial lighting environment, but few concerned about applications without any light resource in night. When used for applications like industrial detection, there is no need to handle scenes without light resource. While for robotics, tasks under poor conditions incapable for humans are usually done without light resources. It could be troublesome for ordinary robotics. Extending the range of the visual sensing and the ability of recognition for robotics, to make them capable for sensing and recognizing scenes in night, and making decisions based on understanding of the scenes, is gradually attracted research attentions in computer vision, both for military and civil applications.Considering the complexity of the scenes for night vision robotics, and the goal to reconstruct panorama of the scenes, a real time depth extraction technology along with a 3-dimentional reconstruction method have been researched. Depth and color information of scenes are extracted simultaneously, to perform a colored reconstruction. With the depth extraction technology, a 3-dimentional model of the night vision scenes could be reconstructed by the computer vision system, which could be used to strength the ability of the computer vision system for understanding and analyzing the night vision scenes. It will be helpful for tasks like object recognition, action recognition, motion planning and automatic navigation.Two distinct methods have been proposed to reconstruct night vision scenes. One is bundle adjustment used widely for daylight scenes reconstruction, the other is based on the depth information of the scenes extracted by near infrared structured light device. Performances of these two methods for reconstructing night vision scenes are evaluated and analyzed.The depth extraction technology used here is based on structured near infrared light. An infrared image pattern is projected to the scene. Distortion of the reflected pattern on the surface could be detected beside the projector. Depth of the surface could be calculated from the distortion, and a depth map is generated.The depth map of the scene could be used to recovery a set of 3-dimentional points in the camera space, which are called point cloud. A few of point clouds recovered from different depth maps captured continuously could be merged by a method called registration after noises and outliers are removed. The merged point cloud is a dense point model of the real world scene.A mesh model of the scene could be constructed by selecting several key points in the dense point model. And the texture of each triangle mesh could be generated by interpolating colors of the vertices. Then a colored mesh model of the real world scene is generated.Experiments show that depth extraction and scene reconstruction technology proposed in this paper could be used to perform a high precise scene reconstruction both in daylight and night vision scenes. The dense point model of the scene could be generated incrementally in real time,and used to produce a mesh model. This technology is effective for the tasks of 3-dimentional reconstruction for night vision scenes.