Research For Scene Depth Structure Of Monocular Image

As a measure of the distance from camera to object in real world, depth provides important 2.5D information for transforming 2D image plane to 3D stereo representation. Monocular depth analysis is to estimate the depth relation between different objects of a scene by leveraging monocular information. As a significant research topic of 3D perception and reconstruction in computer vision, it has wide application in object detection, scene understanding and video surveillance.Depth ordering is an important binary Boolean depth relation, which depicts relative attribute between vision elements in depth. But there are still some difficulties in complex scene. Firstly, occlusion is the major consequence of a projection of the 3D scene onto a 2D image plane. Local depth ordering can be obtained from occlusion relation, but it is difficult to recover occlusion boundary through analyzing 2D image information. Secondly, occlusion cues are sparse and may produce inconsistent global depth ordering in the cluttered scene, so how to eliminate the ambiguity of depth ordering is one of key problem in monocular depth ordering. Thirdly, as reasonable depth ordering should satisfy the constraint of global depth layout, how to construct a depth analysis model, combining local and global depth cues and enforcing global depth reasoning is another problem. Meanwhile, accuracy of object segmentation also affects the quality of depth estimation.To solve these problems, the main work of this dissertation is as follows:(1) For the issue of ambiguity caused by using single occlusion cues in depth ordering. The distance from horizon position to ground-object contact points is used to measure different objects position, thus forming a global depth layout. It provides depth constraints between any two objects, especially between non adjacent regions.(2) Considering local occlusion cues and the global depth cues, CRF model is constructed where object is set as node, the unary potential and binary potential are respectively designed from occlusion cues and global depth layout. In order to obtain more consistent ordering result, belief propagation is leveraged to enforce global depth reasoning by communicating depth information between nodes.(3) To research the effect of motion cues in depth ordering, the motion characteristics of region and boundary are analyzed. We segment the motion objects and judge the occlusion relation between neighboring regions by combining appearance cues and motion cues, which enhance the depth confidence of local region.