Multi-sensor registration is an important and basic problem of intelligent surveillance. A novel visual-thermal image sequence registration method based on motion statistics feature multi-resolution analysis is proposed. In this method, motion statistics feature is utilized to select corresponding point pairs from visual-thermal synchronous video sequence. Then, multi-resolution analysis of motion statistic feature is done to choose proper scale. Finally, outliners are removed by RANSAC, and the geometry transformation parameters are optimized by LM algorithm. By using motion statistics feature, this method avoids the difficult problem of extracting invariant feature from two different image sensor and doesn’t depend on precise motion detection. Through multi-resolution analysis, the proposed approach can resolve the registration under the change of large scale. The performance was demonstrated on three groups of dataset, the results showed that our algorithm carried out precise image registration under the change of translation, scale and rotation.
Learning to See Through a Few Pixels: Multi Streams Network for Extreme Low-Resolution Action Recognition
