Research And Implementation Of Panorama Image Stitching Algorithm

With the development of computer vision technology,the local field of vision images can no longer meet the needs of practical applications.Image stitching is a technology that stitches multiple images that are spatially adjacent and overlap each other into a wide-field image through registration,projection,and fusion.Panorama stitching is a further expansion of image stitching,usually stitching panoramas with a 180-degree or 360-degree field of view.At present,image stitching technology has been widely used in the fields of autonomous driving,virtual reality,medical imaging and drones,but most current image stitching technologies assume that the image is taken by the camera rotating around its optical center or the scene is roughly flat.On the one hand,for complex scenes such as multiple noise and low light,a stable feature detection method needs to be proposed.On the other hand,the current local alignment algorithm improves the alignment capability through complicated calculations,and the problem of time efficiency needs to be solved urgently.In view of the above background,the main research contents and results of this article are as follows:(1)A high-precision feature point detection algorithm HPFA is proposed.Based on the SuperPoint algorithm,a symmetric self-encoder is applied to the feature extraction network,and the feature position information of the image is restored through the hierarchy and incorporated into the scale features of the corresponding layer,which improves the positioning capability of the feature point detection algorithm.At the same time,the self-supervised iterative training method and the traditional manual feature point SIFT are used as auxiliary training data sets to increase the number of feature point detections.Under the evaluation of the Hpatches dataset,HPFA achieves the improvement of feature point repeatability and feature matching ability.(2)A local projection optimization algorithm LPOA is proposed.According to the result of image segmentation,the image is divided into multiple semantically independent regions,and matching is performed by region during the feature point matching process.The region matching method can remove a large number of feature mismatches,and improves the speed of feature matching and global transformation estimation.At the same time,the image is divided into grids for local projection to improve the alignment ability.Similar transformation constraints are introduced to reduce the projection distortion of the image,and the quality of image stitching is significantly improved.(3)Based on the above two image stitching algorithms,a multi-scene image stitching model based on autonomous driving is implemented.Using the CARLA simulator,a method for acquiring panoramic image stitching data was designed for self-driving cars,and straight-line feature alignment constraints were introduced to further enhance its capabilities.Experiments show that the method proposed in this paper can provide a stable wide field of view for autonomous vehicles under different weather and lighting conditions.