使用findHomography和warpPerspective时车牌图像旋转异常问题排查

Hey there! Let's figure out why your warped license plate is rotating 90 degrees—this is almost always an issue with point ordering between your source (plate_polygons) and target (real_plate_polygons) points for the homography transform.

The Root Cause #

cv::findHomography relies on strict, one-to-one correspondence between your source polygon points and target polygon points. If the order (clockwise/counter-clockwise) or the mapping of individual points doesn't match, the transform will twist/rotate the output instead of aligning it correctly.

Looking at your code:

vector< cv::Point> real_plate_polygons; 
real_plate_polygons = {
    cv::Point(PLATE_WIDTH, PLATE_HEIGHT), // 右下
    cv::Point(0, PLATE_HEIGHT),           // 左下
    cv::Point(0, 0),                     // 左上
    cv::Point(PLATE_WIDTH, 0)            // 右上
};

Your target points are ordered bottom-right → bottom-left → top-left → top-right. If your plate_polygons are ordered in a different pattern (e.g., top-left → top-right → bottom-right → bottom-left, a standard clockwise order), this mismatch will cause the 90-degree rotation.

Step-by-Step Fixes #

1. Standardize Point Ordering
   Choose a consistent order (e.g., clockwise starting from the top-left corner) for both source and target points. For a normal horizontal license plate, your target points should look like this:

   vector<cv::Point> real_plate_polygons = {
       cv::Point(0, 0),                  // Top-left
       cv::Point(PLATE_WIDTH, 0),        // Top-right
       cv::Point(PLATE_WIDTH, PLATE_HEIGHT), // Bottom-right
       cv::Point(0, PLATE_HEIGHT)        // Bottom-left
   };
   

2. Ensure plate_polygons Matches the Order
   Verify that the four points in plate_polygons follow the same clockwise/counter-clockwise order as your target points. If your plate_polygons come directly from contour detection, they might be unordered—you'll need to sort them.

   Here's a quick way to sort polygon points clockwise around their center:

   // Calculate the center of the plate polygon
   cv::Point2f center(0, 0);
   for (const auto& p : plate_polygons) {
       center.x += p.x;
       center.y += p.y;
   }
   center.x /= 4;
   center.y /= 4;
   
   // Sort points clockwise based on angle relative to center
   sort(plate_polygons.begin(), plate_polygons.end(), [&](const cv::Point& a, const cv::Point& b) {
       float angleA = atan2(a.y - center.y, a.x - center.x);
       float angleB = atan2(b.y - center.y, b.x - center.x);
       return angleA < angleB;
   });
   

3. Double-Check Dimensions
   Make sure PLATE_WIDTH is larger than PLATE_HEIGHT (since standard license plates are wider than they are tall). If you accidentally swapped these values, it could also lead to a rotated or squashed output.

Quick Verification #

Before running the transform, draw the ordered plate_polygons on your source image using cv::polylines—this will let you confirm the points are arranged in the expected clockwise/counter-clockwise order.

Once you align the point order between source and target, the warpPerspective output should be correctly oriented, just like your other working examples!

内容的提问来源于stack exchange，提问作者Jose Tomas