| Image change detection has a wide range of applications in computer vision and remote sensing,such as anomaly detection,infrastructure inspection,disaster monitoring and prevention using remote sensing satellite images or surveillance cameras,etc.However,most of the current change detection algorithms are based on the unit of image pixels,and require accurate calibration to compare the pixel changes between images taken at different times and different cameras.The difference of viewpoints between image pairs taken by satellites or surveillance cameras is small and easy to calibrate.However,image calibration is difficult in the case of moving camera position,and such algorithms are less robust to viewpoint changes.In addition,pixelbased change detection methods are susceptible to noise conditions such as illumination changes,and the detected objects usually do not include moving objects in the world coordinates.At the same time,for moving objects,most of the current motion detection and tracking algorithms are based on video or image sequence,which need to continuously monitor the screen,and need to extra and extract multiple video frames or comparison.Therefore,in order to solve the above problems,this paper carries out research on the change detection of multi-view image pairs based on moving objects.Instead of using the traditional pixel as the unit,the change detection based on moving objects is carried out at the object level.The main results are as follows:(1)A change detection algorithm for moving objects in multi-view images based on epipolar constraint is proposed.Firstly,ASIFT algorithm was used to detect image feature points,and RANSAC algorithm was used to eliminate mismatching feature point pairs.Secondly,the epipolar constraint is introduced and combined with image features.Finally,the change detection of moving objects in images taken at any different time and from different angles is realized by using a robust and efficient template matching algorithm based on deep convolutional features.The running time of the algorithm can meet the real-time requirements,and the accuracy rate of the algorithm is 68%.It meets the change detection needs of moving targets in the new application scene of images with large changes in shooting perspective,which is of great significance in actual production and life.(2)A multi-view image-pair object-level change detection network OBJ-CD-MVIMONet based on moving objects was proposed.The network is studied for image pairs with viewpoint differences.Different from previous algorithms,this paper treats this change detection task with viewpoint differences as an object matching problem of two images extracted from each image.Therefore,the proposed network is able to detect object-level changes with viewpoint differences more accurately than existing pixellevel methods.Moreover,unlike the vast majority of existing algorithms,the network does not require pixel-wise change annotations.The network proposed in this paper uses the object detection module MASK R-CNN to extract the objects in each image,and then uses the Siamese matching network module based on image similarity evaluation and the Transformer-based COTR graph matching network module to compose the object matching module to construct the correspondence between objects and make a judgment on the correspondence between objects.This is used to detect the appearance or disappearance of objects in the scene and the movement of objects.Experimental results show that the accuracy of the proposed network is better than that of the existing change detection algorithms,and it is robust to the change of viewpoint,and can adapt to object-level change detection with large changes of viewpoint.At the same time,it is the first network that includes the object with changed position,that is,the moving object,into the change detection range.(3)The data sets MAD-OBJ-CD,CARLA-OBJD and SD-OBJ required by OBJCD-MVI-MONet object level change detection network are constructed.MAD-OBJCD is an object-level change detection dataset generated by CARLA simulator to simulate real-world scene changes.It contains 1000 image pairs with large view difference and labels three types of changed objects: appearance or disappearance objects and moving objects with changed ppositions CARLA-OBJD is a data set for object detection in CARLA scene,which contains 1000 images and annotations of four types of objects: cars,buildings,street lights,and pedestrians.The SD-OBJ dataset contains more than 50,000 images and is mainly used to improve the similarity measure ability for the above four categories of objects. |
