Feature Point

• Basic concepts
  • Definition: The point of local uniqueness in the frame
    • It should be able to discern even if the object’s size, position, camera point of view, lighting, etc. change.
• Applicable fields: recognition, detection, tracking, classification…
• property
  • Repeatability: The same area can be found in different frame.
  • Differentiation: Feature points in different region have different expressions.
  • Efficiency: The number of feature points in the image should be smaller than the number of pixels.
  • Locality: Feature points relate only to the local area of the image.
• Configuration(구성)
  • Keypoint: Location information of feature points
  • Descriptor: A vector describing information about surrounding pixels
• Harris corner detector
  • In a small window in an image, the part with large image change in all directions is regarded as a corner
  • The point where the image change value of the pixel is locally maximized is used as the corner point.

• Shi-Tomasi
  • Good features should be able to be optimized for tracking algorithms!
  • Both simple translation and affine changes are considered.
  • Only the smallest of the two eigenvalues is considered.

Scale Invariant Feature Transform(SIFT)

• You need strong points such as changes in size, direction and lighting!
  • Result of picking feature vectors for local patches centering on identifiable feature points
  • It expresses the direction and rapid degree of regional brightness change.
  • Due to the large amount of computation, it is not suitable for real-time SLAM implementation.
    • Of course, it can be complemented by GPU programming
  • The Difference of Gaussian method was used to compensate for the large amount of computation in the Laplacian of Gaussian method.
  • Laplacian: 2nd derivative of the brightness of the image

• Difference of Gaussian(DoG)

• Process
  • Scale-space extrema detection: Potential feature point candidates robust to changes in size and direction are extracted based on the difference of Gaussian for all scales.
  • keypoint localization: For each extrema, low-contrast points and points in the edge region are removed.
  • Orientation assignment: directional assignment based on local gradient for each point
  • Keypoint descriptor: 16x16 size block -> 4x4 sub-block unit
    • Generate 8 bin histogram of direction in each sub-block
    • Finally, create a 128 dimensional vector

Oriented-FAST and Rotated-BRIEF(ORB)

• quickly detect feature points based on the apparent change in brightness of grayscale

• Binary Robust Independent Elementary Features(BRIEF)
  • Descriptor vector is composed of binary numbers.
  • Decide the binary number to treat as the vector through comparison between two pixels in the patch

• Random two pixel pairs are randomly determined based on probability distribution
• Generate 128bit binary descriptor by comparing 128 pairs

• Oriented-FAST
  • The weakness of FAST is that it does not have direction and scale.
  • Solve scale immutability problems with image pyramids
  • Connect the geometric center O and the moment-based center C to find the direction for it
• Rotated-BRIEF
  • Conventional BRIEF does not have rotational invariance
  • Use of keypoint direction information obtained during FAST detection
  • Use Steer BRIEF: Performs BRIEF after rotating according to the direction information

Correspondence matching(대응 매칭)

• Basic concepts
  • Process required to perform data association process in SLAM
  • The process of correctly matching descriptors between images and images, or images and maps
  • Problems caused by repetitive textures can cause inconsistent accumulation!
• Brute-force matcher
  • The point with the shortest descriptor distance between feature points in two different images is regarded as a matching point.
  • Generally, descriptor distance means similarity
  • Euclid distance, Hamming distance…
  • If the number of dots increases, the calculation amount increases, which is not efficient.
• Fast Library for Approximate Nearest Neighbors(FLANN)
  • Matching library for fast matching in high Dimension
  • Real-time matching is suitable for SLAM

Reference

SLAM KR