Study On Moving Object Detection Algorithm With The Moving Camera

In the field of computer vision,as the basic work of many advanced vision tasks,moving object detection is widely used in the fields of intelligent security,intelligent transportation,national defense and military.At present,traditional moving object detection algorithms are mostly aimed at scenes where the camera is in a stationary state.However,with the rapid development of mobile computing platforms in recent years,more and more video data are captured by freely moving cameras,such as handheld camera,pan-tilt camera,and vehicle-mounted camera.In the scene where the camera can move freely,the background in the video moves at all times,and the moving object is no longer the only factor that causes changes between frames.The mixed motion of the background and the foreground makes the task of moving objects detection more difficult.In this case,how to distinguish the respective motion patterns of the foreground and the background and extract the real motion object from the complex background motion accurately and effectively has become a new problem to be solved urgently.This paper mainly focuses on the research of the moving object detection algorithm with the moving camera.The main work of this paper is as follows:(1)Focuses on the modeling method of background motion between adjacent frames of video and the constraints of the establishment of the model,and explains in detail how to use the background motion model to compensate for background motion between adjacent frames,so as to realize the extension of traditional background modeling algorithms to the scene where the camera can move.(2)Under the simple background motion scene that satisfies the homography constraint condition,due to the limitation of the accuracy of the background motion model estimation,registration errors between pixels usually occur when the background motion between frames is compensated.Existing algorithms based on background modeling usually use a fixed neighborhood background model search method to reduce the impact of pixel registration errors,so as to ensure the accuracy of detection.But in this process,if the algorithm uses an inappropriate search area size,it will cause a serious decline in the recall rate of the detection.In response to the above problems,this paper designs an adaptive neighborhood background model search method,which uses the reprojection error of pixels to adaptively determine the size of the search area corresponding to each pixel.This method not only reduces the impact of registration error,but also effectively solves the drawbacks caused by the fixed neighborhood search method.Experimental results show that,compared with the fixed neighborhood search method,the adaptive neighborhood search method can well balance the accuracy and recall rate,reduce the "holes" that appear on the foreground objects in the detection results,and effectively improve completeness of target detection results.(3)Under the complex background motion scene that does not meet the homography constraints,because the background motion between frames can no longer be described by the homography matrix,the algorithm based on background modeling will fail.In addition,most of the existing moving object detection algorithms for complex background motion scenes are offline algorithms and cannot be applied to online detection scenes.In response to the above problems,this paper designs an online moving object detection algorithm based on multiple segmentation.This algorithm decomposes the moving object detection task into the intersection of motion segmentation and general object segmentation.Firstly,according to the dense optical flow between sequential frames and the motion inconsistent boundary in the optical flow amplitude field and direction field,the appropriate foreground and background seed points are automatically selected.Secondly,based on the selected seed points,the visual optical flow field can be segmented to the background/foreground area by the OneCut segmentation algorithm.Finally,the result of motion segmentation is combined with the result of general object segmentation to complete the final motion object detection.In order to solve the problem of general object segmentation or motion segmentation failure caused by target deformation and inconsistent motion boundaries in some video frames,we design a forward propagation algorithm to optimize the motion segmentation and general object segmentation results of each frame.Experimental results show that the proposed algorithm can accurately extract the moving object from the complex background without any hypothetical constraints on the camera motion,and its performance is superior to other same kinds algorithms.