Motion Target Detection Of RGBD Image Based On Scene Flow

With the advancement and development of technology,computer vision is more and more widely used in people's lives.Through the powerful computing power of computers,key information existing in images can be extracted and utilized.The detection of moving targets is one of the hottest research direction.Moving object detection usually analyzes the motion of each object in the image plane by analyzing the change of the pixel position of the corresponding object in the video or continuous image sequence at different times.At present,moving target detection technologies in medical,transportation,military and other fields play an important role.At present,the moving target detection technology can be divided into two categories,one is the static target detection based on the background subtraction method and the frame difference method,and the other is the dynamic target detection based on the optical flow method.The scene flow used in this paper is a moving object detection technology based on RGBD image,which is a three-dimensional instantaneous velocity solution method obtained by the optical flow method through the constraint of adding depth information.With the improvement of the technology of a number of depth sensors such as Kinect,the acquisition of depth information is no longer dependent on the parallax ranging method of multi-view cameras.The estimation of scene flow is also developed from the original based on multi-view camera parallax calculation method to based on RGBD Image scene flow technology with the popularity of consumer-grade depth sensors,RGBD scene flow method has also become a hot topic of research.This paper mainly uses the scene flow method to detect moving objects in the RGBD image sequence,focusing on the difference between the moving target and the background in the scene.The main work of this paper is to apply the scene flow estimation method to the detection of moving objects and optimize part of the process of scene flow calculation.Firstly,in the acquisition process of RGBD images,because the structured light technology used by the depth sensor can not accurately capture all the information in the scene,some key position information may be missing in the depth image.In this paper,the depth image is filled with holes using the similarity around the pixel points before the depth image is used,which better restores the depth information of the key points.In order to avoid the influence of the global motion caused by camera and sensor motion,this paper uses the camera motion model based on Fast feature point matching to correct the coordinates of corresponding pixel points between two frames of RGBD images before and after camera motion.The corrected coordinate points are used to estimate the scene flow.In the final experimental results,the method of this paper solves the influence of global motion on the detection result.In the previous scene flow estimation,in order to detect large motion in the image,the method of feature point matching is adopted.As long as the feature point is in the image before and after the motion,there is no mismatch.However,this method has certain drawbacks,it can not be calculated in the case of too many or too few feature points,and the obtained scene flow is also sparse.The dense scene flow needs to be further calculated according to the scene structure.In this paper,the multi-scale iterative calculation method is used to reduce the image size of large motion until it can be directly calculated,and further iterate to large size,finally realize the detection of large displacement motion.Aiming at the phenomenon that the moving target and the background boundary are blurred in the optical flow target detection method,this paper adopts the strategy of combining global smoothing and local smoothing to increase the gradient at the boundary,and the detection effect on the moving target edge is better in the experiment.