OpenCV图像鸟瞰图异常排查：为何出现多余白色区域？

Hey there! Let's dig into why you're getting those unwanted white regions in your bird's-eye view, and fix the issues step by step.

1. Critical Typo in Perspective Transform Source Coordinates #

First off, there's a clear bug in your source vertex definition that's messing up the perspective warp:

src_vertices[2] = Point(upper_righty, upper_righty);

You accidentally used upper_righty for both the x and y coordinates here—this should be upper_rightx for the x-value! That mistake distorts the source quadrilateral you're trying to warp, so areas outside your intended lane region are being included in the bird's-eye view, leading to extra white spots.

Fix it to:

src_vertices[2] = Point(upper_rightx, upper_righty);

2. Overly Aggressive Preprocessing (Histogram Equalization + Thresholding) #

Your current preprocessing pipeline (histogram equalization followed by binary thresholding) is likely enhancing non-lane white pixels:

• Global histogram equalization boosts contrast across the entire image, which can turn faint pavement stains, shadow edges, or other light-colored road features into bright white pixels.
• If your binary threshold is too low, it'll capture all these unintended bright pixels along with the lane lines.

Better Preprocessing Fixes: #

• Use color filtering first: Lane lines are almost always white or yellow—filter for these colors before grayscale/equalization to eliminate most non-lane pixels. For white lanes, use HSV color space to isolate high-value, low-saturation pixels:

  Mat hsv_img;
  cvtColor(original_img, hsv_img, COLOR_BGR2HSV);
  
  // Define white color range (tweak these values for your specific lighting)
  Scalar lower_white = Scalar(0, 0, 200);
  Scalar upper_white = Scalar(180, 30, 255);
  
  Mat white_mask;
  inRange(hsv_img, lower_white, upper_white, white_mask);
  

• Swap global equalization for CLAHE: Adaptive histogram equalization (CLAHE) enhances contrast locally without over-amplifying noise or faint background details:

  Mat gray_img;
  cvtColor(original_img, gray_img, COLOR_BGR2GRAY);
  
  Ptr<CLAHE> clahe = createCLAHE(2.0, Size(8, 8)); // Adjust clip limit and tile size as needed
  clahe->apply(gray_img, gray_img);
  

• Combine mask + thresholding: Apply the white mask to your CLAHE-enhanced grayscale image before thresholding to only keep pixels that are both white and bright:

  Mat filtered_img;
  bitwise_and(gray_img, white_mask, filtered_img);
  
  Mat binary_img;
  // Use Otsu's automatic thresholding to find the optimal cutoff
  threshold(filtered_img, binary_img, 0, 255, THRESH_BINARY | THRESH_OTSU);
  

3. Ambiguous Output Size in getBirdView #

Your getBirdView function uses dst.size() for the output resolution, but if dst isn't pre-initialized to 640x480 before calling the function, this can lead to unexpected warping or scaling. Explicitly define the output size to match your target coordinates:

void getBirdView(Point2f *p1, Point2f *p2, const Mat& src, Mat& dst) {
    Mat warpMatrix = getPerspectiveTransform(p1, p2);
    // Hardcode the target resolution to match your desired 640x480 bird's-eye view
    warpPerspective(src, dst, warpMatrix, Size(640, 480), INTER_LINEAR, BORDER_CONSTANT);
}

Final Notes #

Start with fixing the coordinate typo—it's the most immediate issue. Then refine your preprocessing pipeline to isolate only white lane pixels before applying the perspective warp. With these changes, you should see only the two parallel lane lines in your bird's-eye view.

内容的提问来源于stack exchange，提问作者yooo123