Stereo Visual Odometry Robust Against Blur Images

This thesis focuses on how to reduce the impact of blur images on stereo visual odometer and improve odometer’s positioning accuracy in real-time. The challenges are:1. How to estimate image’s blur degree in real time.2. How to classify images accurately while the scene varies.3. How to deal with images of different blur degrees so that the positioning accuracy of stereo visual odometry can be improvedThe main contributions of this thesis are as follows:1. According to characteristics of blur images and latent images, a real-time image blur evaluation algorithm based on distribution of small gradient is put forward. The method computes image’s blur degree in real-time, which is the foundation of image classification.2. According to images cameras taken, combining inertial filters with FIR filters, a video blur grade classification system is proposed. The classification system computes image’s blur degree by the real-time image blur evaluation algorithm, then classify images in real time according to image’s blur degree.3. A stereo visual odometry method robust against blur images is put forward. The method classifies images using the video blur grade classification system. Different computing strategy will be used to different image level. The proprosed method is tested and verified on different visual odometry platforms. The proprosed method is also compared to others, which proves that the proprosed method is robust and can be applied widely.