Research And Application On Gigapixel Imaging Algorithm For Multi-view And Cross-resolution Images

For the past few years,more and more application scenarios have rely on ultra-high pixel images and videos.However,traditional single camera imaging is increasingly unable to meet people’s demand for big views,such as giant Image Maxium screens,and omni-directional imaging equipment for vehicles.The scope of vision detection in the conventional monitoring system is small,and it is impossible to take into account the big field of vision(FOV)and high-definition details.The accurate tracking of dense objects is still challenging.Therefore,Camera Arrays based on multi-camera imaging is proposed to spliced and restructured cross-resolution images through image registration,and ultra-high pixel image imaging of large fields is performed.The use of non-structured camera arrays to expand the detection range while retaining local clear details is of practical significance for obtaining more comprehensive tracking information.The feature point extraction and matching of the image plays a key role in solving the camera arrays image synthesis.Due to the huge resolution differences in multi-cameras,the collection images have large visual spots.Low the algorithm distribution accuracy will cause problems such as pseudo-shadows and distortion in synthetic images.The existing image registration algorithm has a large amount of calculation and low accuracy,especially for cross-resolution images poorly.The higher positioning accuracy and the characteristics of the characteristic point of the individual estimation ability detection and describe the network,the efficient cross-resolution image synthesis method,and the more accurate long-view small target recognition and tracking algorithm are all issues that need to be solved by the current research.In summary,the following work has been carried out in this article:(1)A new network that ACPoint is proposed to solve the low accuracy of cross-resolution image matching and large positioning error of existing feature point extraction algorithms.The ACPoint combines the precision of multi-branch network and the reasoning speed of flat network structure.Meanwhile,a feature-point labels dataset is created based on transfer learning,and the model training and iteration are carried out in a self-supervised manner.In addition,a model adaptation technique is proposed to filter labels by comparing and verifying the labels of datasets.Experiments show that the proposed network model has better homophily estimation ability and positioning accuracy.(2)A new image composition algorithm based on unstructured camera arrays is proposed to solve the problem of small FOV of conventional camera.In this algorithm,multi-scale and multi-field images are used to composition large-field gigapixel images.Through multi-level feature matching and image warping,the long focal view is embedded into the short focal view to provide the details of the local view,and the color transfer algorithm is used to maintain the color consistency between different views.The main merits of this algorithm are that it can solve the huge resolution gap,the angle difference and the color inconsistency between the images.Experimental results show that the proposed image composition scheme can generate a large field of view image at the level of gigapixels without the constraint of camera calibration and local view overlap area.(3)In order to solve the low recognition rate of small targets in realistic scenarios,a new targets dataset in road environment was reconstructed,and the YOLOv5s(You Only Look Once)target detection model was improved for retraining.Based on the above image composition algorithm,an unstructured camera arrays has used to generate high-definition composite video for urban road scenes,and vehicles and pedestrians are detected and tracked based on the composite video.Deep Sort object tracking algorithm was used to track the trajectory of people,cars,trucks and buses.Experiments show that compared with traditional video,the YOLO＿STD target detection model combined with Deep Sort algorithm proposed can improve the effect of multi-object detection and tracking.