Paper reading note-CMT tracking algorithm

First, the author's thesis consensus-based Matching and Tracking of keypoints for Object Tracking:

1 in the first frame, the feature points of the target are found with the brisk feature descriptor, each of which is represented by relative coordinates relative to the center of the target region;

2 starting from the second frame, each frame is then used brisk descriptors to find the feature points, in order to match the first frame target's feature points, each feature point and the first frame of each feature point of the Euclidean distance, and the nearest and second nearest to the ratio as the scale to determine the feature point and the first frame which target feature point most match, The index of the feature point is recorded after the successful match;

3 and 2 parallel, using the optical flow method, the target feature point of the previous frame is used to predict the feature points of the current frame, and the index values of these feature points are unchanged.

4 Synthesis 2 and 3, take the same set, the same feature point index is inconsistent, the 2 is the same, after Fusion records these feature points in the image of the absolute coordinate values;

5 The current frame Target Center coordinates are obtained by subtracting the relative coordinate value of the first frame from the absolute coordinate value of the current frame. In order to solve the target area scaling and angular rotation, the scaling factor and rotation matrix are multiplied by the relative coordinates of the feature point at the first frame when the difference is made.

6 The central coordinates of each key point may be inconsistent, and we think most are right, so with the voting (clustering) mechanism, the position with the highest number of votes is the central position.

7 after getting the center position, let's find the position of Four corners in the target area by multiplying the relative coordinates of the corner of the target area by the center coordinate + the first frame by the scaling factor and rotation matrix.

Note: Both the rotation angle and the zoom factor are selected with the median

The reason for choosing brisk features is that the feature has rotation and scale invariance, and is a binary representation, convenient to calculate Euclidean distance.

The author then published the Clusteringof Adaptive correspondences for deformable object tracking in 15 cvpr,

is a complement to the CMT algorithm, first of all, the static global search is the 2nd step above the minimum minimum ratio before the limit and add a minimum distance constraint (equation 1); then the main contribution of the paper, a non-similarity measurement method (equation 2), the specific method is: to any two key points, In the first frame corresponding to the key point after the rotation scaling transformation (the transformation factor is H) after the transformation of the coordinates, the transformation of the coordinates are poor, the difference is less than the set threshold value (the threshold is 20), the cluster to the target group, otherwise the cluster is a non-target group; In order to eliminate the ambiguity in the key point matching process of the previous paper (for example, the right eye of the current frame may match the left eye of the initial frame), a round of key matching is performed, but at this point the initial frame is not retrieved globally, but only the key points in the initial frame are taken into account in the target group.

Note: In the non-similarity measurement method, take two points to subtract, because the tracking method, the coordinates of the first frame is relative coordinates (relative to the target center), and then each frame with absolute coordinates, two points subtraction, in order to subtract the difference between the relative and absolute coordinates, and two points into a line, relative to the single point of transformation, After the difference between two points, it is better to resist the effect of the target displacement rotation.