A Survey on Real-Time Motion Estimation Techniques for Underwater Robots

Author: Ferreira

Year: 2014

Notes:

Feature Detection:

Template extraction, extraction of high spatial wavelength and high variance templates
Shi-Tomasi = improvement of Harris corner detector based on hessian matrix
SIFT based on difference of the image convolved with different gaussian kernel
SURF similar to SIFT with higher robustness to scale change
FAST if at least n contiguous pixels in a circle are darker than a threshold (Accelerated Segment Test), it is considered as a feature best for n=9
Laplacien $L(x,y) = \frac{\partial^2 I(x,y)}{\partial x^2} + \frac{\partial^2 I(x,y)}{\partial y^2}$
STAR computation of Laplacian of Gaussian in a star shape => avoid subsampling + non maxima suppression
ORB like fast but with an orientation operator
AGAST like FAST but with an adaptative decision tree
BRISK like FAST but performed at different scale

Descriptors:

For template, the normalized correlation is used between the patches
SIFT based on the gradient orientation histogram in a subregion of 4x4 pixels
SURF based on Haar wavelet response in the neighbourhood, and nearest neighbour approach
BRIEF binary descriptor that are compared with a serie of XOR operation = computationnaly efficient
ORB = rotated brief, improvement of brief that takes roation into account
BRISK = rotated BRIEF with a method based on orientation computation with a sampling of points
FREAK = binary, bio-inspired : like brisk but the sampling is more concentrated in the center, coarse to fine search

Motion Estimation:

Front end of a real time SLAM with feature tracker and laser + acoustic altimeters
Gives the depth of pixel Z
Vehicle model based on constant heading and altitude
Estimation of the speed with a least square approach and motion model

Experiments: