| ... | ... | @@ -97,6 +97,97 @@ With $I$ as the intensity in grayscale at the point with coordinates (x,y). The |
|
|
|
Low point density: Number of points $\in \lbrack 300\ ;500\rbrack$
|
|
|
|
Medium point density: Number of points $\in \lbrack 1000\ ;2000\rbrack$
|
|
|
|
High point density: Number of points $\in \lbrack 3000\ ;6000\rbrack$
|
|
|
|
The density can be chosen via the Image Stabilization Parameters menu, and it determines the required computation time for image stabilization.
|
|
|
|
|
|
|
|
Next, each point is associated with a 128-element descriptor vector, which depends on the pixel values within a neighborhood of 8 to 16 pixels. We try to match the descriptors of the $N$ points between two images and eliminate false matches. The 25% of points with the most deviating descriptors are eliminated. The SURF function is launched by the image stabilization function in Fudaa-LSPIV with the following command: |
|
|
\ No newline at end of file |
|
|
|
To ensure that stabilization is performed on the banks, the user defines a "flow" area in which no points of interest will be extracted.
|
|
|
|
|
|
|
|
{width="700px"}
|
|
|
|
|
|
|
|
## :mag: Point Matching
|
|
|
|
|
|
|
|
The next step is to match the points from image 1 with those from image 2. We then use the SURF descriptor, which is essentially the "identity card" of each point. It contains several pieces of information:
|
|
|
|
|
|
|
|
- Point scale (its "size")
|
|
|
|
- Orientation
|
|
|
|
- Distribution of gradients in its neighborhood
|
|
|
|
|
|
|
|
The neighborhood is defined based on the size of the point, and the gradients are expressed relative to its orientation. As a result, the SURF descriptor is robust to rotation, translation, and zoom. Points with similar SURF descriptors are then matched. In the image below, the matches are indicated by a red line. For each point, the diameter of the circle represents the size of the point of interest, and the line indicates its orientation.
|
|
|
|
|
|
|
|
{width="400px"}
|
|
|
|
|
|
|
|
Sometimes incorrect matches (or false matches) are made. To discard them, we use the ORSA method (Moisan 2004). False matches are identified using a stochastic process associated with epipolar geometry constraints (not detailed here).
|
|
|
|
|
|
|
|
## :mag: Estimating the Transformation
|
|
|
|
|
|
|
|
We now have sets of matched points, meaning we know the position of each point on image 1 and image 2. The next step is to determine the transformation matrix $H$ that allows us to go from positions $X$ to $X'$.
|
|
|
|
|
|
|
|
```math
|
|
|
|
X=H \cdot X^{\prime}=\left(\begin{array}{lll}
|
|
|
|
h_1 & h_2 & h_3 \\
|
|
|
|
h_4 & h_5 & h_6 \\
|
|
|
|
h_7 & h_8 & h_9
|
|
|
|
\end{array}\right) \cdot\left(\begin{array}{c}
|
|
|
|
x^{\prime} \\
|
|
|
|
y^{\prime} \\
|
|
|
|
1
|
|
|
|
\end{array}\right)
|
|
|
|
```
|
|
|
|
|
|
|
|
Each matched point is used as a control point. The values of the homography matrix coefficients $H$ are then approximated using them. The DLT (Direct Linear Transform) method is used to rearrange the above equation and estimate $H$ using the pairs of control points. Indeed, there is no unique solution to the problem, so the best-fitting solution is determined for all the control points.
|
|
|
|
|
|
|
|
:clap: **End of deep dive** :nerd_face:
|
|
|
|
|
|
|
|
#
|
|
|
|
|
|
|
|
</details>
|
|
|
|
|
|
|
|
The **_Images 🡪 Image Stabilization Settings_** menu provides access to the stabilization module.
|
|
|
|
|
|
|
|
{width="600px"}
|
|
|
|
|
|
|
|
In this panel, the **_Enable the stabilization module_** checkbox activates or deactivates the stabilization of the source images.
|
|
|
|
|
|
|
|
 
|
|
|
|
|
|
|
|
Next, it is necessary to define the extent of one or more **_Flow Zones_** or **_Fixed Zones_** by clicking the **_Interactive Input_** button and entering the positions of the polygon vertices on a source image. A flow zone corresponds to moving patterns that should not be used as points of interest by the stabilization algorithm. Conversely, a fixed zone includes fixed patterns that can be used as points of interest by the stabilization algorithm. A **_Flow Zone_** layer is added to the 2D view **_Source Image Space_**.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
By default, the point of interest density is **_Medium_**, and the transformation model is **_Similarity_**. Two models are proposed: **_Similarity_** and **_Perspective_**. The choice depends on the magnitude of camera movements and the "quality" of the banks.
|
|
|
|
|
|
|
|
- Small movements (<10 pixels): prefer the **_Similarity_** model.
|
|
|
|
- Large movements (>10 pixels): prefer the **_Perspective_** model.
|
|
|
|
|
|
|
|
- Poor "quality": choose **_High_** point density.
|
|
|
|
- Good "quality": choose **_Low_** point density.
|
|
|
|
|
|
|
|
> The **"quality"** of the banks refers mainly to **texture and contrast**. The more textured and contrasting the environment, with visible edges/corners, the better the bank will be for stabilization.
|
|
|
|
> - A smooth concrete slab without roughness will be seen as a poor "quality" bank.
|
|
|
|
> - Banks made of vegetation + rocks + trees + shadows will be of very high quality.
|
|
|
|
|
|
|
|
The **point density** affects the **calculation time**: the lower the density, the faster the stabilization process.
|
|
|
|
|
|
|
|
> The banks/flow ratio in the image is an important parameter to consider.
|
|
|
|
> The test case presented earlier is ideal: **the flow represents about 1/3 of the image**, leaving enough banks. The environment is also sufficiently **textured** and has many edges. In such a case, a **_Low_** density can be used.
|
|
|
|
>
|
|
|
|
> It is not recommended to use **Low** density with the **Perspective** model. This model is more sensitive, so errors can occur if there are too few control points.
|
|
|
|
To ensure proper stabilization, **the banks should represent at least ¼ of the image**, on both sides of the flow (i.e., 50% flow, 50% banks).
|
|
|
|
|
|
|
|
Finally, you must **_Validate_** or **_Cancel_** the defined settings. The **_Apply_** button has no function here. The stabilization settings are reported in the gauging reports.
|
|
|
|
|
|
|
|
The **_Images 🡪 Stabilize Images_** command launches the calculation of the stabilized images, which will then be used for orthorectification (see the next section). The stabilization process can be interrupted.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The stabilized images can be viewed one by one in the 2D View **_Source Image Space_** by first clicking on the "Shaky Hand" icon .
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The **_File 🡪 Export 🡪 Export Stabilized Images_** command allows the export of stabilized images.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#
|
|
|
|
-> [**Next Section: 5. Image Orthorectification**](User-Manual/Image-Orthorectification) |
