Real-time Video Image Mosaic Research And Technical Implementation

It is normal to use a wide view angle in Intelligent Transportation for tracking moving objects, but a single camera often can't meet the requirements, thus it is necessary to use realtime video image mosaic technology by overlapping required scenes with multiple cameras. Real-time image mosaic is a hot topic in recent years, it can be applied to remote sensing image, intelligent transportation, security monitoring, medical imaging, and many other fields. This paper is aimed at getting real time stitching from different camera angles, then providing a wide picture perspective of real-time monitoring. This paper is intended to ensure that the system can get the real-time performance requirement under the premise of precision image mosaic effect, then achieve the optimal value both in integral and local performances.Scale invariant feature extraction operator is a common feature extraction method in image mosaic, but the computing speed of traditional scale invariant feature extraction operator is very slow. In order to meet real-time requirements, this paper should improve the speed and accuracy of feature extraction in order to obtain a superior adaptability transition model. Therefore, this paper proposes a ROI feature extraction method based on SURF, and uses image-based perceptual hashing algorithm dynamic template in matching method to estimate the area of interest, this algorithm can effectively reduce the features detection range of image and improve detection accuracy. RANSAC is often used in feature matching algorithm to wipe off wrong matching points, but the matching accuracy of traditional RANSAC is limited to the selection of the feature point, and it is easy to have some mismatches. In order to improve the matching precision, the paper put forward three improved ways for RANSAC algorithm: the first one is to set thresholds based on the Euclidean distance, then remove some wrong feature points; the second is one to estimate the ROI, and use the information of ROI as priori knowledge to determine the interior point RANSAC algorithm standard; the third one is to modify the standard of determining the optimal model in RANSAC, we propose a optimum sample selection based on confidence of variance. The improved matching algorithm can get a better performance in scale, lighting, rotation and stretching and other aspects.In order to meet the real-time requirements, in the re-mapping and image fusion part, this paper firstly calculate the best suture, then achieve fast fusion in circumscribed rectangular area, comparing with conventional method of calculation in the overlap region directly, this algorithm can better meet the real-time requirement. The improved algorithm in this paper can achieve the average speed of 12.3 frames per second without hardware acceleration when using video images of a resolution of 640 * 480, it can meet the real-time performance, and ensure optimum splicing results.