Vehicle localization based on scene perception

The vision-based vehicle localization task is tackled as an image-based place recognition problem in this thesis. As image representation is an important process to place recognition, we proposed place recognition methods focusing on increasing the quality of image features. First, the dynamic object removal step is proposed to remove the dynamic objects, such as vehicles and pedestrians, of an image by semantic segmentation method and restore their background information by image inpainting method. Second, reducing instead of removing the noisy information of an image is proposed. Using the image blurring method to reduce the image noise, the performance of place recognition is improved wi…