Research On Fish-eye Image Panoramic Stitching Method Based On DSP

Fisheye lens has many advantages such as the large field of view and short focal length.Constructing a panoramic vision system to use the fisheye lens, can get the panoramic images use the the single visual system by one shot. It has been widely used in the security surveillance, robot navigation, aerospace and other fields. Although the fisheye lens image has the advantages of wide field of view, there are large distortion, it will affect the panoramic image stitching effect. This paper studys the image matching technology according to the characteristics of fisheye lens.The steps of the fisheye image panoramic stitching mainly includes image acquisition,the distortion correction, image registration and image fusion.The distortion of the collecting fisheye images are large. Firstly, it needs to correct the distortion images to complete the stitching. For the drawback of traditional fisheye image correction method, such as spherical coordinate orientation and latitude and longitude method,the correction effect is very poor. This paper proposes a fisheye lenses distortion correction method based on the spherical perspective projection constraint. According to the perspective projection model, determin the distortion parameters, and the correction effect is better.However, if correcting the entire fisheye image, correction effect will poor, the phenomenon of stretching is serious, will bring some difficulties to stitch. If the two images has the overlap region, the correction range can be appropriately reduced to improve the quality of image stitching. This paper proposes an improved method to quantify fish-eye image and determine the correction range and meet the requirements of stitching at last.It can be able to achieve a better matching effect based on Scale Invariant Feature image registration techniques. However, a fisheye image is large, the running time of SIFT matching algorithm is long and there are a lot of false matches. In view of the shortcoming of SIFT algorithm, this paper proposed a feature point detection method based on D2 OG operator,which effectively reduce the running time in matching. RANSAC algorithm is used to eliminate false matches and ensure the matching accuracy.Because shooting error and light, the image obtained after stitching, will have some problems such as ghosting and sutures. The linear fusion method can smooth the transition and finally get a better effect panoramic image.Finally, the image stitching algorithm was transplanted into the embedded panoramic vision system based on DSP and the algorithm is optimized, which achieves the desired requirements.