Pixel correspondences for SLAM using sidescan sonar with canonical representations

Sammanfattning: Acoustic instruments play an important role in autonomous underwater vehicles (AUVs). Sidescan sonar (SSS) detects a wider range, provides photo-realistic images in high resolution, and has lower price than the multibeam echo sounder (MBES). However, SSS projects the 3D seafloor to the 2D image, which are distorted by the AUV’s altitude, target’s range and sensor’s resolution. In the pixel correspondence tasks, an area is expected to be as similar as possible in SSS images from different survey lines. In this thesis, a canonical transformation method is proposed to decrease the above distortion. The transformation includes reducing the intensity decay effect of incident angle with Lambertian cos, cos2, cot laws, and projecting the bin’s position into equally spacing horizontal range. A dataset with pixel correspondence annotation is built as ground truth in this thesis. Based on this dataset, two experiments on similarity measure of patch pairs and descriptor matching with canonical SSS images are presented. The results show that canonical transformation can improve the performance of descriptor matching as well as the feature’s similarity in different images.