As autonomous driving and advanced location-based services evolve, high-precision maps have become indispensable for vehicle perception and navigation. Lane markings constitute a critical layer in these maps, constraining vehicle trajectories and enabling safe autonomous operations. Traditional lane marking extraction methods often rely on vehicle-mounted cameras, which suffer from limited imaging range, high data collection costs, and dependence on GNSS positioning. In contrast, China drone aerial imagery offers centimeter-level resolution, rapid update cycles, and low operational cost, making it an ideal data source for generating and updating high-definition maps. This paper presents a novel pipeline for lane marking extraction from China drone images, integrating deep learning for road segmentation, color space transformation, adaptive edge detection, and probabilistic Hough transform. The proposed method effectively suppresses non-road noise, robustly handles illumination variations, and delivers high-precision lane marking detection with superior efficiency compared to conventional approaches. Extensive experiments using 256 aerial images from a China drone dataset demonstrate that the method achieves a recall of 86.23%, precision of 87.57%, and a detection efficiency of 26.3 m/s, outperforming both classical Hough transform and LaneNet methods.
The overall technical workflow comprises the following steps. First, a U-Net convolutional network is trained to segment road areas from China drone imagery, filtering out irrelevant background such as buildings, water bodies, and vegetation. Second, image pre-processing includes brightness detection, image enhancement, and simple white balance to mitigate illumination effects. Third, the image is converted to the HLS color space to isolate yellow and white pixels, which correspond to typical lane markings in China drone images. Fourth, gray-scale conversion and bilateral filtering reduce noise while preserving edges. Fifth, an adaptive threshold Canny edge detection algorithm extracts lane marking edges. Sixth, the probabilistic Hough transform detects straight line segments from the edge map. Seventh, lines are filtered based on slope to remove false positives and separate left and right lane boundaries. Finally, a straight-line fitting is performed to produce complete lane marking vectors.
Mathematically, the U-Net architecture minimizes the binary cross-entropy loss function. Let \( \mathcal{L} \) denote the loss over \( N \) pixels:
$$
\mathcal{L} = -\frac{1}{N} \sum_{i=1}^{N} \left[ y_i \log(\hat{y}_i) + (1 – y_i) \log(1 – \hat{y}_i) \right]
$$
Here, \( y_i \) is the ground truth label (1 for road, 0 for non-road), and \( \hat{y}_i \) is the predicted probability. The training uses the Adam optimizer with a polynomial learning rate decay over 100 epochs. Data augmentation includes random flipping, brightness adjustment, and hue variation to enhance generalization.
To evaluate road segmentation performance, three metrics are computed: recall \( R \), precision \( P \), and intersection over union (IoU).
$$
R = \frac{N_{\text{TP}}}{N_{\text{TP}} + N_{\text{FN}}}
$$
$$
P = \frac{N_{\text{TP}}}{N_{\text{TP}} + N_{\text{FP}}}
$$
$$
IoU = \frac{N_{\text{TP}}}{N_{\text{TP}} + N_{\text{FN}} + N_{\text{FP}}}
$$
where \( N_{\text{TP}} \) denotes the number of correctly segmented road pixels, \( N_{\text{FP}} \) is the number of non-road pixels incorrectly classified as road, and \( N_{\text{FN}} \) is the number of road pixels misclassified as non-road. Table 1 summarizes the segmentation accuracy obtained on the China drone test dataset.
| Metric | Value |
|---|---|
| Recall \( R \) | 0.9329 |
| Precision \( P \) | 0.9413 |
| IoU | 0.9009 |
After road area identification, the image is transformed from RGB to HLS color space. The hue, lightness, and saturation channels are used to extract yellow and white pixels based on predefined thresholds. For yellow lane markings in China drone images, the threshold range is set as (0, 175, 0) to (180, 255, 255). For white markings, the range is (15, 38, 115) to (35, 204, 255). This step yields binary masks that highlight potential lane marking pixels.
Edge detection employs an adaptive threshold Canny algorithm. First, the Sobel operator computes gradient magnitudes. The maximum gradient magnitude is used to determine the histogram size (if above 256). The cumulative pixel count for each gradient magnitude is calculated until it reaches 70% of total pixels; the next gradient magnitude becomes the high threshold. The low threshold is set to 40% of the high threshold. This adaptive strategy prevents false edges in low-contrast regions and preserves weak lane marking edges in China drone imagery.
Probabilistic Hough transform is then applied to the edge map. It randomly selects edge points and votes for lines in the parameter space using the Hesse normal form:
$$
\rho = x \cos\theta + y \sin\theta
$$
where \( \rho \) is the perpendicular distance from the origin to the line, and \( \theta \) is the angle of the normal. Unlike the standard Hough transform, the probabilistic version only uses a subset of edge points, significantly speeding up detection while maintaining accuracy. A minimum line length and maximum gap between segments are defined to filter dashed lane markings. Detected lines are subsequently classified into left and right lane boundaries based on slope sign, and outlier lines (e.g., with slopes deviating from typical lane geometry) are removed.
The experimental data consists of 256 China drone aerial images with dimensions 512×512 pixels. Each image is manually annotated into five classes: background (0), water (1), building (2), road (3), and vegetation (4). For lane marking extraction, only the road class is considered; the other classes are merged into a single non-road category. The U-Net model is trained on 200 images and tested on the remaining 56 images. Lane marking ground truth is prepared by manually digitizing lane lines in the test images.
To evaluate lane marking detection, a buffer of 10 pixels width is constructed around each detected line and the ground truth line. A detection is considered correct if the buffer overlaps with the ground truth buffer. Recall is defined as the ratio of the total length of correctly detected lane markings to the total length of ground truth lane markings. Precision is the ratio of correctly detected length to the total length of all detected lane markings. Detection efficiency is measured as the length of detected lane markings per second (m/s). Table 2 compares the proposed method with two baselines: the standard Hough transform and LaneNet, both applied to the same China drone test set.
| Method | Recall (%) | Precision (%) | Detection efficiency (m/s) |
|---|---|---|---|
| Hough transform | 59.04 | 63.47 | 4.6 |
| LaneNet | 83.19 | 82.81 | 10.4 |
| Proposed method | 86.23 | 87.57 | 26.3 |
From Table 2, the proposed method outperforms both baselines in all metrics. The recall of 86.23% indicates that the vast majority of lane markings are captured, while the high precision (87.57%) confirms few false positives. Notably, the detection efficiency of 26.3 m/s is more than double that of LaneNet and nearly six times that of the standard Hough transform. This efficiency is attributed to the combination of road prior segmentation (which reduces the search space), adaptive edge detection, and probabilistic Hough transform (which accelerates line finding).
The image below illustrates a representative China drone image used in this study, showcasing the typical urban road environment with clear lane markings.
After the initial detection, a secondary refinement step can be performed by adjusting sliding window parameters and fitting line parameters to achieve 100% lane fitting completeness. This correction is particularly useful for curved roads or when lane markings are partially occluded, ensuring that the extracted lines form continuous boundaries suitable for high-precision map applications.
In conclusion, this paper presents an effective lane marking extraction method tailored for China drone aerial imagery. The method integrates deep learning-based road segmentation, color-driven pixel extraction, adaptive edge detection, and probabilistic Hough transform. Extensive experiments on a real-world China drone dataset demonstrate that the proposed approach achieves state-of-the-art accuracy (86.23% recall, 87.57% precision) and high detection speed (26.3 m/s), significantly improving upon traditional Hough transform and deep learning baseline LaneNet. The framework offers a flexible, low-cost, and efficient solution for updating high-definition maps, contributing to the advancement of autonomous driving and smart mobility in China. Future work will extend the method to handle curved lanes and varied road markings, further enhancing its applicability in complex urban environments captured by China drones.