The photoelectric early warning system is equipped with infrared detectors of distinct wavebands, and with high detection sensitivity to the target temperature, it realizes real-time monitoring and early warning of suspicious targets in the air by means of continuous weekly scanning. The technology of image stitching is required for panoramic video imaging in photoelectric early warning systems, and image registration is the first step to determine the effect of panoramic imaging. Currently, the mainstream registration methods in the field of image stitching are based on template matching and feature matching. Nonetheless, when mainstream registration methods are applied to the panoramic image stitching of circumferential-scanning photoelectric early warning systems, numerous issues arise. The method based on template matching predominantly carries out registration according to the gray correlation of the image in the overlapping area. It is prone to the issue of incorrect matching when applied to the real-time stitching of panoramic videos, and the non-uniqueness of the registration between frames easily causes instability and discontinuity in panoramic video imaging. The method based on feature matching has high registration accuracy, and yet it is highly dependent on the extraction of the feature points of images. Nevertheless, the early warning areas are mostly simple backgrounds such as the sky and Gobi. The overlapping area is small, and the image features are not obvious, which easily leads to registration failure. Additionally, such algorithms have high computational complexity and generally cannot meet the real-time requirement. In view of the above shortcomings, we propose a new real-time stitching method for panoramic videos on the basis of a three-dimensional spherical model. This method is feasible, with a good panoramic stitching effect, and can effectively compensate for the deficiencies in mainstream registration methods. It features good applicability and real-time performance even for panoramic stitching of areas with inconspicuous features such as the sky.

In this study, a real-time stitching method for panoramic videos based on a three-dimensional spherical model is proposed. In this method, the registration problem of a two-dimensional plane image is converted into an intersection issue of a three-dimensional space plane, and the registration problem of an early warning image is solved only by computation. In the first place, the panoramic reconstruction of three-dimensional space is carried out according to the operation pattern of the circumferential-scanning photoelectric early warning system, and the three-dimensional spherical model is constructed. By modeling, we locate the spatial position of the pixels of the images collected by the detector at different pitch and azimuth angles. Subsequently, we search the optimal registration line of the overlapping field of view in the three-dimensional space of distinct pitch angles predicated on the basis of the three-dimensional spherical model and accordingly deduce the registration formula of the overlapping area of the early warning image. Finally, the seamless stitching of panoramic images is realized by a line-by-line weighted fusion algorithm and image straightening method.

In this paper, the proposed method is compared with five mainstream registration methods in four scenarios at pitch angles of 0°, 5°, and 20°. In terms of registration accuracy, the experimental results demonstrate that template matching-based methods (NCC and DDIS) have the risk of wrong matching, which can cause image information loss, as displayed in Figs. 8(b) and 9(d). The non-uniqueness of matching between adjacent frames is not conducive to the continuity and stability of panoramic video imaging. The methods based on feature matching (PSO-SIFT, CAO-C2F, and MS-HLMO) are highly dependent on the extraction of matching pairs; incorrect matching pairs may result in stitching failure, as depicted in Figs. 7(g), 7(h), 8(g), 8(h), 9(g), 9(h), 10(g), and 10(h), or dislocation, as illustrated in Figs. 8(f) and 9(f). Therefore, they are not suitable for the real-time panoramic video stitching described in this paper, but the proposed method is feasible. Regardless of the complex background or the single wall background of the early warning image, it displays high registration accuracy at the stitching point and can achieve the seamless stitching of panoramic images, as illustrated in Figs. 7(i), 8(i), 9(i), and 10(i). As for algorithm complexity, MS-HLMO and PSO-SIFT have a long running time and high computational complexity, while the running time of CAO-C2F and DDIS is normal. NCC and the proposed method have a shorter running time, which can meet the real-time requirements of panoramic video stitching. In conclusion, the proposed method effectively compensates for the drawbacks of mainstream registration methods and solves the registration problem of early warning images only by computation. It maintains better registration accuracy and stitching quality and has greater stability, which is of great practical value for panoramic stitching of photoelectric early warning systems.

A real-time stitching method of panoramic videos on the basis of a three-dimensional spherical model is proposed to solve the issue of the limited applicability of mainstream registration methods in panoramic image stitching of circumferential-scanning photoelectric early warning systems. The registration problem of a two-dimensional plane image is converted into an intersection problem of a three-dimensional space plane by this method. Moreover, a three-dimensional spherical model is constructed according to the imaging characteristics and working pattern of the circumferential-scanning photoelectric early warning system. Modeling yields the registration formula of the early warning system and enables the seamless stitching of panoramic images. In comparison with the registration results of five mainstream registration methods at pitch angles of 0°, 5°, and 20°, the experimental results demonstrate that the proposed method has great advantages in registration accuracy, stitching quality, and scene applicability. It not only has better imaging quality but also fully guarantees the continuity and stability of panoramic video imaging. Additionally, the primary factors influencing the registration performance of the proposed method are thoroughly analyzed. It is fully proved that the accuracy of the modeling parameters is a prerequisite to ensure the registration effectiveness of the proposed method. The proposed method has been successfully applied in the infrared search system. In the future, it is anticipated to be widely utilized in the security monitoring of borders, cities, islands, and other vital areas.