| Visual SLAM(Visual Simultaneous Localization and Mapping)is a process in which a mobile carrier carries a visual sensor to locate itself through the motion process and simultaneously construct a map of the surrounding environment.The key of VSLAM is to estimate the pose of camera.Visual Odometry(VO),as the front end of VSLAM,is an effective method to estimate the pose of camera.The visual sensor used by VO is usually visible light camera.However,visible light camera is susceptible to the influence of lighting conditions,especially in the environment of poor illumination,dynamic illumination or smoke,so the VO accuracy of visible light camera is poor.Compared with visible image,infrared imaging is not affected by light.But the lack of detail texture and contrast in infrared images often makes them less effective than visible light cameras in well-lit environments.This thesis combines the characteristics of the two kinds of images to study the visible-infrared fused visual odometry which can adapt to different lighting conditions.The research content of this thesis is as follows:1.Considering the differences between the two different images,a fusion strategy is used in this thesis.First,the infrared image describing the same scene is processed with the visible image.Discrete Wavelet Transform,based on discrete Wavelet Transform image fusion scheme,then highlight the hot objects found in the infrared image in the visible image by monochromatic threshold image fusion technology,Dynamically adjust the fusion parameters that best suit your environment.In the evaluation of fusion performance,the classical monocular VO algorithm is used to test on public datasets with different images as input.The results show that the VO accuracy of fusion images is higher than that of visible and infrared images,and the error is less than 2.61%.2.For the above fusion images,a visual odometry method based on deep learning is used in this thesis.The algorithm trains the deep neural network to first learn and extract the global features of inter-frame changes in monocular fusion images,and then learn to conduct sequence modeling for the extracted features,and output the relative inter-frame pose of the camera.By comparison with the representative visible visual odometry and infrared thermal odometry,the positioning accuracy is better in different lighting scenes,which is improved compared with the traditional monocular VO algorithm,and the error is less than 2.41%.In order to verify its generalization performance,experiments across data sets are also carried out,and the accuracy remains good under the new scenario.In summary,the fusion strategy used in this thesis combines the advantages of infrared and visible images and achieves good performance on the traditional visual odometry,which verifies that visible infrared image fusion is of great significance for low-light and dynamic lighting scenes.On this basis,with the fusion image as the input,the deep learning-based visible-infrared fused visual odometry designed in this thesis can adapt to different lighting scenes,and through the experiment of different data sets,it is verified that the algorithm has high precision and good generalization ability. |
