基于HOG描述符训练线性SVM人体检测分类器时出现训练错误

It looks like you’re hitting snags with your HOG-SVM setup, and there are a few key issues in your current code that we can sort out step by step:

Common Problems in Your Approach #

• Incorrect Feature Matrix Shape: You initialized x as a (3780, 2) array, but HOG descriptors are 1D vectors (typically thousands of elements long). Your feature matrix should follow the shape (number_of_samples, descriptor_length) instead.
• Incomplete HOG Computation: Your code cuts off before calculating the HOG features for your samples—this is critical because the SVM relies on these descriptors to learn patterns.
• Single Sample Limitation: Training an SVM with just one positive and one negative sample won’t yield a useful detector. Classifiers need diverse examples to generalize, but let’s first fix the code to work with your test setup.

Corrected Code Example #

Here’s a complete, working version that properly computes HOG descriptors and trains a Linear SVM:

import numpy as np
import cv2
from sklearn.svm import LinearSVC

# Initialize HOG descriptor with standard human detection parameters
hog = cv2.HOGDescriptor(
    winSize=(64, 128),
    blockSize=(16, 16),
    blockStride=(8, 8),
    cellSize=(8, 8),
    nbins=9
)

# Load and preprocess positive/negative samples
positive_img = cv2.imread('G:/Project/Database/db/positive_sample.jpg', cv2.IMREAD_GRAYSCALE)
negative_img = cv2.imread('G:/Project/Database/db/negative_sample.jpg', cv2.IMREAD_GRAYSCALE)

# Resize images to match HOG window size (critical for consistent descriptor length)
positive_img = cv2.resize(positive_img, (64, 128))
negative_img = cv2.resize(negative_img, (64, 128))

# Compute flattened HOG descriptors for each sample
hog_positive = hog.compute(positive_img).flatten()
hog_negative = hog.compute(negative_img).flatten()

# Build feature matrix and label array
X = np.vstack([hog_positive, hog_negative])
y = np.array([1, 0])  # 1 = human, 0 = non-human

# Train the Linear SVM
svm = LinearSVC()
svm.fit(X, y)

# Optional: Test the classifier with a sample image
test_img = cv2.imread('test_human.jpg', cv2.IMREAD_GRAYSCALE)
test_hog = hog.compute(cv2.resize(test_img, (64, 128))).flatten()
prediction = svm.predict(test_hog.reshape(1, -1))
print(f"Prediction: {'Human detected' if prediction[0] == 1 else 'No human'}")

Key Fixes Explained #

1. HOG Parameter Setup: We use industry-standard parameters for pedestrian detection (64x128 window size, which aligns with common human dataset dimensions).
2. Image Resizing: All samples must match the HOG window size to ensure the descriptor length is consistent across inputs.
3. Feature Matrix Construction: Flattened HOG descriptors are stacked into a 2D array where each row represents one sample.
4. Clear Labeling: Explicit labels for positive and negative samples are required for supervised training.

Critical Tips for Reliable Results #

• Use More Samples: Training with two samples will not produce a functional detector. The Pascal human dataset has hundreds of samples—aim for at least 50 positive and 100 negative examples.
• Data Augmentation: For positive samples, add rotations, flips, or small shifts to increase diversity and improve generalization.
• Hard Negative Mining: After initial training, use the classifier to find false positives in negative images and add those to your training set to refine accuracy.

内容的提问来源于stack exchange，提问作者Chetan Birajdar