Concept of Image Processing

For binocular vision, I think the difference is in image processing. According to my current understanding, one aspect of binocular vision is camera calibration, and the other is the matching of feature points. Since no specific tests have been conducted, the understanding of many things remains to be improved.

For camera calibration, we now want to use a very mature method to directly use the camera calibration toolbox in MATLAB.

Therefore, the difficulty lies in image processing, that is, searching for feature points for matching.

Think about it. What we need to identify now is very simple, and the goal is also very clear, if you do not consider special circumstances. That is to say, a small ball with a color near Red and a yellow ball at both ends, with a blue angle column in the middle, and the target is not small. However, the difficulty is that the image captured by USB camera is captured during motion and jitters occur, so that the image will be blurred. Now I want to summarize the problems found during image processing in the previous two days, and then I want to find a solution.

There are many ways to separate images. Color-based and shape-based. Now I feel that although the problem does not seem very difficult, it would be impossible to use a single image segmentation method.

A few days ago, we always considered only color-based division. That is to say, the RGB color space is converted to the HSI color space and then combined with the S value. However, it is mainly divided according to the H value. In this exampleProgramSet the threshold, because in general sense, these are still very clear colors. In this way, there will be some messy points that do not actually belong to the areas I want to be split up, and some points that are clearly in the areas I want, not separated. For vertices that are separated but not targets, you can use opening and closing in morphology to remove them. However, in this way, the targets will also suffer some losses. When we used the color method, we wanted to obtain the coordinates of each vertex in the target, and then obtain the value of the center. According to this processing effect, the obtained center point will certainly have a large deviation. For binocular vision, feature point matching is the most important. Therefore, this is definitely not acceptable. We have also considered using the regional growth method, but the regional growth method should take into account the selection of seed points, the growth conditions and the termination conditions. The growth conditions here should be the most difficult. The key is that I have not compiled similarAlgorithmBut now there is no time to consider this experiment. In addition, I also thought that using this method will still lead to problems where some areas cannot be separated. In general, we cannot use the color division method!

Now let's think about it. Since it requires a higher accuracy, it should be better to use edge detection. It may be better to use color image edge detection. This morning, I tried to use common Sobel and other operators to detect blurred images because of camera movements and edges (here, it is first converted to grayscale images). Finally, almost none of the edges have been detected. If you are using the Canny operator, there will be a lot of things to be detected. The moderation operator is indeed good, but there are too many details. In this way, what you really want will be drowned. In the above experiment, the grayscale images are used. In this way, for the characteristics of angle columns and small balls, they are curved and smooth. Therefore, the effects of brightness and light on them are quite large, when some light intensity occurs, some RGB values are near the same. In this case, the result is that when the HSI model is converted, the S value is near 0. In this way, the H value does not have much meaning (in this way, my color-based splitting effect is not very good ). The other influence is that the contour is smooth, so that edge detection based on the gray value is not effective (except for the Canny operator ).

Now, I want to combine color-based segmentation and edge detection. Finally, I want to detect the straight line and circle (ball) with the use of the Hough transform ). It seems that the effect is not very good because it is used separately. Maybe this can be achieved through Edge Detection of color images. However, first of all, I have not found any good methods for Edge Detection of color images, I have searched for this content, but there is nothing to gain. I have read the image processing book that says that for edge detection, color images can be the same as gray images, that is to say, the three RGB components are done separately, which is also better than the HSI effect. Secondly, I am still thinking that if we use edge detection, we need to use the end of the line-of-the-art transformation. If the end of the line-of-the-art transformation is very large, then first, there will be a lot of unwanted interference, and it will increase the amount of computing. Therefore, we would like to use color-based segmentation to determine the area to be split first, and then perform edge detection and Hough Transformation in this area. In this way, the color-based and shape-based transformations are combined.

Edge detection is currently in trouble. If you want to try the algorithm on Matlab, as long as the algorithm is good, the program, I think, should not be too difficult. Then, the final program will be compiled in the evening or tomorrow. Tomorrow is week 4.

The key is the algorithm, for fear that a good algorithm cannot be found. Image processing is so important to the entire binocular vision. If the image processing is poor, then the entire system will definitely not work.

Yes. Take one step to ensure that the maximum possible errors of one step are not too many.