Binocular Visual Tracking Model Incorporating Inertial Prior Data

The binocular vision tracking model is an important part of binocular visual inertial positioning and is a prerequisite for determining the accuracy of visual inertial positioning and system robustness. In the typical visual inertial positioning, the tracking model is mainly divided into two processes, namely feature matching and pose tracking model. Feature matching is a key process in visual positioning. It mainly solves the problem of image data association between adjacent frames in visual positioning. The traditional visual feature matching firstly uses the corresponding feature matching algorithm to perform a rough matching on the feature points of adjacent frames, and then uses the random consistent sampling algorithm to reject all the matching points by mistake matching. However, these algorithms all have different degrees of error, and the random sampling consensus algorithm requires iterative data iteration, which greatly increases the system matching time. Based on this, we propose a culling mismatching method that combines inertial prior data. Experiments show that the feature matching model based on inertial prior data is much better than the feature matching model without prior data, which indicates that the feature matching method based on inertial prior information is feasible.