cv2.HoughLinesPointSet在简单测试案例中检测结果不佳的技术求助

Let's break down why your test case isn't working as expected and fix it step by step. Your setup is straightforward (two perfect horizontal lines), so the issue almost certainly lies in parameter configuration and minor edge-case handling.

What's Going Wrong? #

Your current parameters are too loose, allowing noisy, low-vote lines to clutter your results, while the actual target lines might not be prioritized. Let's go through the key issues:

1. threshold=5 is too low: Each of your lines has 10 points, so a threshold of 5 lets lines supported by just half a line's worth of points slip through.
2. rho_step=1 is too coarse: While your lines align exactly with rho=0 and rho=1, larger step sizes can lead to vote dispersion if there's even tiny numerical error.
3. No handling for near-vertical angles: Horizontal lines correspond to theta=np.pi/2 (90 degrees), where tan(theta) becomes infinite—your current code handles theta=0 but not this edge case, leading to potential numerical instability.
4. lines_max=20 returns too many noise lines: You only need 2 lines, so this parameter is overkill and floods your output with irrelevant results.

Fixed Code with Explanations #

Here's the adjusted code with comments explaining each change:

import matplotlib.pyplot as plt
import numpy as np
import cv2 as cv2

# Test case (unchanged)
x = np.linspace(0, 10, 10)
y1 = np.zeros(10)
y2 = np.ones(10)
plt.scatter(x, y1)
plt.scatter(x, y2)

# Prepare point set (unchanged)
pnts1 = np.column_stack((x, y1))
pnts2 = np.column_stack((x, y2))
pnts = np.row_stack((pnts2, pnts1))
pc = pnts.reshape(-1, 1, 2).astype(np.float32)

# Updated Hough parameters
pc_lines = cv2.HoughLinesPointSet(
    pc,
    lines_max=2,  # Only return the top 2 lines (we know there are 2)
    threshold=9,  # Require at least 9 votes (close to our 10-point line count)
    min_rho=0,
    max_rho=20,
    rho_step=0.1,  # Smaller step for finer rho resolution
    min_theta=0,
    max_theta=np.pi,
    theta_step=np.pi/180  # 1-degree step for precise angle detection
)

# Extract results
votes, rho, theta = pc_lines[:, 0][:, 0], pc_lines[:, 0][:, 1], pc_lines[:, 0][:, 2]

# Plot with robust line conversion
xx = np.linspace(0, 10)
for (vote, r, t) in zip(votes, rho, theta):
    print(f"Detected line: Votes={int(vote)}, Rho={r:.2f}, Theta={t:.2f} rad")
    
    # Handle horizontal lines (theta ~ pi/2)
    if np.abs(t - np.pi/2) < 1e-3:
        yy = np.full_like(xx, r)
    # Handle vertical lines (theta ~ 0 or pi)
    elif np.abs(t) < 1e-3 or np.abs(t - np.pi) < 1e-3:
        plt.axvline(x=r, color='orange')
        continue
    # Regular line equation
    else:
        a = -np.cos(t) / np.sin(t)  # More stable than 1/tan(t)
        b = r / np.sin(t)
        yy = a * xx + b
    
    plt.plot(xx, yy, label=f"Votes: {int(vote)}")

plt.legend()
plt.show()

Key Improvements #

• Tighter threshold: Ensures only lines with strong support (almost all points on the line) are detected.
• Finer resolution steps: Smaller rho_step and theta_step make it more likely the exact parameters of your horizontal lines are captured.
• Edge-case handling: Directly computes horizontal/vertical lines instead of relying on unstable trigonometric calculations.
• Limited line count: lines_max=2 cuts through noise and returns only the lines you care about.

Additional Notes for Real-World Use #

If you're working with noisy point clouds later:

• Lower the threshold to around 60-70% of the expected points per line.
• Consider preprocessing your point cloud (e.g., removing outliers) before running the Hough transform.
• You can sort the detected lines by vote count to prioritize the most robust ones.

内容的提问来源于stack exchange，提问作者yonatan