Target Bounding Box Adjustment Method Based on Feature Matching and Improved Three-channel RANSAC Filtering

In common target tracking algorithms, the focus is often on the location of the target, rather than on some of the changes of the target its own. Especially when the size and rotation of the target change, the bounding box is difficult to accurately adapt to this change. In this regard, we use SIFT feature matching method to find the corresponding feature points on the front and back frame targets, and get the mean value according to the distance and angle changes of all matching points as the basis of adjusting the bounding box. However, this method has a higher requirement for the accuracy of feature matching, otherwise, once there are some large deviations in the matching results, it will have a great impact on the final adjustment. In order to ensure reliability, besides improving the matching accuracy, a large number of accurate results are needed to calculate as well. Therefore, we adopt a matching screening method that combines three-channel similarity and RANSAC, which select a large number of accurate matching results as the basis to calculate the change of size ratio and rotation angle, to ensure that the adjustment of bounding box is accurate and reliable.