Research For Depth Topological Structure Of Monocular Image Based On Occlusion Detection And Scene Layout

The depth in vision system is to measure the distance between the observed object and camera equipment.It is an important 2.5D information representation that transforms the 2D vision information acquired from the sensor device to the 3D vision.In computer vision,monocular depth structure can be defined as figure-background relationships of objects in monocular image,in general,this monocular depth relationships can only shows that object is farther or closer to the camera,cannot obtain the absolutely depth value of objects.According to Gestalt's psychological theory,this depth perception method is consistent with human's visual perception of the world,which is to build a reasonable and consistent depth layout structure.Depth ordering is a relative foreground-background relationship of objects in the scene,but in complex and chaos scenes,the depth relationship between the objects is not easily to obtained,the main reasons are:firstly,in complex scenes,it is difficult to obtain accurate segmentation and the segmentation quality directly affects the results of depth ordering;secondly,occlusion is the main consequence of a projection from 3D world to 2D plane,which can be used to infer the local depth relationship,however,it is difficult to detect the occlusion relations between regions based on 2D image;thirdly,occlusion relation is relatively sparse and may produce inconsistent depth ordering,in addition,the occlusion relation obtained may be in conflict with the real world,and how to eliminate the ambiguity of depth ordering is very important;fourth,a reasonable depth ordering must satisfy the constrain condition of global depth layout,thus,how to construct a global depth model,which combines multiple depth cues to enforce depth ordering reasoning is a difficult.To solve those problems,this paper carried out the following works:(1)To solve the problem of occlusion in complex scenes,we construct a deep convolution network framework to learn the occlusion relationship,and we can obtain depth ordering based on occlusion relationships.(2)For the issue of ambiguity depth ordering caused by using only a local occlusion cues,vanishing point is considered to analyzed in our depth ordering model,which is a benchmark to estimate the distance between vanishing point and ground contact points,then form a global depth structure layout to estimate depth of arbitrary objects,providing a constraint of depth ordering.(3)Considering local occlusion cues and global depth cues obtained by ground contact points,graph model is constructed with objects as node,and belief propagation is leveraged to enforce global depth reasoning based on the local occlusion relationship and global depth layout.(4)Combining depth relationships between regions and appearance characteristics of super-pixel regions,a global energy optimization framework is constructed,and a hybrid evolution algorithm is utilized to minimize the energy function to obtain more consistent depth ordering results.