In recent years, the rapid advancement of autonomous driving and high‑precision mapping has placed increasing demands on the accuracy and efficiency of lane marking extraction. Traditional approaches rely heavily on vehicle‑mounted cameras, which suffer from limited field of view, high data collection costs, and dependency on GPS signals. To address these bottlenecks, I have turned to drone technology as a flexible and cost‑effective alternative. In this work, I present a novel pipeline for extracting lane markings from high‑resolution drone imagery, integrating deep learning semantic segmentation, color space analysis, adaptive edge detection, and probabilistic Hough transform. The proposed method achieves superior precision and recall while maintaining real‑time processing capabilities. The extensive use of drone technology not only reduces the cost of high‑definition map updates but also enables large‑scale, frequent monitoring of road infrastructure.
1. Introduction
The emergence of autonomous vehicles has created an urgent need for high‑definition maps that contain detailed lane‑level information. Lane markings are critical elements that regulate vehicle trajectories and ensure safe navigation. Historically, lane detection systems have been developed using images captured by onboard cameras, followed by convolutional neural networks (e.g., LaneNet) or feature‑based methods. However, vehicle‑based data acquisition presents several drawbacks: (1) high operational cost limits the update frequency of high‑definition maps; (2) reliance on GNSS positioning leads to information loss in signal‑degraded environments; (3) perspective transformation from the camera view to a bird’s‑eye view introduces projection errors; (4) the imaging range is inherently narrow, making large‑scale detection inefficient. Drone technology offers a promising solution by providing centimeter‑level resolution, wide coverage, and rapid deployment capability. Despite its potential, research on lane marking extraction from drone imagery remains sparse. Existing methods, such as Aerial LaneNet, attempt to directly segment lane markings from aerial views, but they struggle with background noise from non‑road areas and the small size of lane targets. In this paper, I propose a comprehensive framework that leverages drone technology to overcome these challenges. The core contributions are: (1) employing a U‑Net network for accurate road area segmentation to filter out irrelevant background; (2) applying HLS color space transformation and adaptive Canny edge detection to isolate lane marking pixels; (3) utilizing probabilistic Hough transform for robust line detection; and (4) integrating slope‑based filtering and line fitting to produce complete lane models. Experiments on real drone datasets demonstrate that my method outperforms traditional Hough transform and LaneNet methods in both accuracy and speed.
2. Overall Technical Pipeline
The entire workflow is designed to process raw drone images and output clean lane markings. The steps are as follows:
- Road area segmentation using a trained U‑Net model.
- Image brightness detection followed by image enhancement and white balance correction.
- Conversion to HLS color space and extraction of yellow and white pixel masks.
- Grayscale conversion and bilateral filtering to reduce noise.
- Adaptive threshold Canny edge detection.
- Probabilistic Hough transform for line segment detection.
- Slope‑based filtering to eliminate false positives and separate left/right lanes.
- Linear fitting of the detected lane segments.
The pipeline fully exploits drone technology by leveraging the high spatial resolution and orthorectified nature of aerial views, which eliminate the need for perspective transformation and reduce systematic errors.
3. Methodology
3.1 U‑Net for Road Area Segmentation
I adopt the U‑Net architecture, a classic semantic segmentation model originally designed for biomedical images. Its U‑shaped encoder‑decoder structure effectively captures multi‑scale features. The encoder consists of repeated convolutional and max‑pooling layers that gradually reduce spatial dimensions while increasing feature channels. The decoder uses transposed convolutions for up‑sampling and skip connections to fuse low‑level and high‑level features, enabling precise pixel‑wise classification. In this work, I train U‑Net to distinguish road pixels from non‑road pixels (background, buildings, water, vegetation). The training dataset includes 256 aerial images of size 512×512, manually labeled into five classes (background: 0, water: 1, building: 2, road: 3, vegetation: 4). For the lane extraction task, I binarize the labels into road (class 3) and non‑road (all others). Data augmentation techniques such as random flipping, brightness adjustment, and hue variation are applied to improve generalization. The model is trained for 100 epochs using the Adam optimizer with a polynomial learning rate decay and a binary cross‑entropy loss. The loss descent curve shows stable convergence. Evaluation metrics include recall (R), precision (P), and intersection‑over‑union (IoU):
$$ R = \frac{N_{TP}}{N_{TP} + N_{FN}} $$
$$ P = \frac{N_{TP}}{N_{TP} + N_{FP}} $$
$$ IoU = \frac{N_{TP}}{N_{TP} + N_{FN} + N_{FP}} $$
where NTP is the number of correctly segmented road pixels, NFP is false positive, and NFN is false negative. The achieved results are summarized in Table 1.
Table 1: Road area segmentation accuracy
| Metric | Value |
|---|---|
| Recall (R) | 0.9329 |
| Precision (P) | 0.9413 |
| IoU | 0.9009 |
The high IoU (over 90%) confirms that U‑Net effectively isolates road areas, significantly reducing the influence of non‑road texture in subsequent lane detection.
3.2 Image Preprocessing and Color Space Conversion
Drone imagery is subject to varying illumination conditions. I first apply a simple white balance algorithm to correct color casts. The principle is to equalize the R, G, B values for pixels that should appear white, thereby establishing a proper color mapping across the image. Next, I convert the image to the HLS (Hue, Lightness, Saturation) color space, which is robust to illumination changes and facilitates the separation of yellow and white lane markings. Through empirical tuning, I define the following thresholds: yellow lane pixels are extracted in the range (0, 175, 0) to (180, 255, 255) in HLS; white lane pixels are extracted in the range (15, 38, 115) to (35, 204, 255). After thresholding, I combine the yellow and white masks to form a binary image containing only potential lane marking pixels. This step leverages the inherent color properties of lane markings and dramatically reduces false positives from road surface texture.
3.3 Adaptive Threshold Canny Edge Detection
The classic Canny edge detector is sensitive to the choice of high and low thresholds. To automate this process, I design an adaptive thresholding strategy. First, I compute the gradient magnitude using the Sobel operator. If the maximum gradient value exceeds 256, I set the histogram bin size accordingly. Then, I accumulate the pixel counts for each gradient level until the cumulative sum reaches 70% of the total pixels; the gradient level at that point is selected as the high threshold. The low threshold is set to 40% of the high threshold. This method effectively adapts to the local contrast and noise characteristics of drone images. The output of adaptive Canny is a binary edge map that highlights the boundaries of lane markings while suppressing irrelevant edges.
3.4 Probabilistic Hough Transform for Line Detection
Standard Hough transform is computationally expensive and tends to produce many duplicate line segments. I employ the probabilistic Hough transform (PHT), which randomly samples a subset of edge points and votes in the parameter space. PHT returns the endpoints of detected line segments, which directly represent the lane markings. The algorithm parameters include the pixel resolution (typically 1 pixel), angular resolution (1 degree), and a minimum number of votes threshold. To further refine the results, I filter detected lines based on their slopes. In a typical road scene, lane markings appear as near‑vertical or slightly slanted lines. Segments with slopes outside a reasonable range (e.g., horizontal lines) are discarded. The remaining segments are separated into left and right lane groups according to their positions and slopes. Finally, I perform linear least‑squares fitting on each group to generate continuous lane models. The combination of adaptive Canny and probabilistic Hough transform is a key advantage of my approach, as it balances detection accuracy and computational efficiency, fully utilizing the capabilities of drone technology for real‑time applications.
4. Experiments and Results
4.1 Dataset
The experimental dataset consists of 256 drone‑acquired aerial images with a size of 512×512 pixels. The images cover various road types (straight, curved, intersections) under different lighting conditions. The dataset is divided into training (80%) and testing (20%) sets. Ground truth lane markings are annotated manually. The high spatial resolution (centimeter level) provided by drone technology allows precise labeling of lane boundaries.
4.2 Lane Detection Performance
I evaluate the proposed method using recall and precision, defined as the ratio of correctly detected lane pixels to ground truth lane pixels (recall) and the ratio of correctly detected lane pixels to total detected lane pixels (precision). A buffer of 10 pixels around the ground truth lane centerline is used to tolerate minor deviations. In addition, I compare against two baseline methods: (1) standard Hough transform combined with fixed‑threshold Canny edge detection, and (2) LaneNet, a deep learning‑based lane segmentation network. The comparison also includes detection efficiency, measured in meters per second (m/s), in terms of lane length processed per unit time. The results are summarized in Table 2.
Table 2: Performance comparison of different lane extraction methods
| Method | Recall (%) | Precision (%) | Detection Efficiency (m/s) |
|---|---|---|---|
| Standard 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, my method achieves the highest recall (86.23%) and precision (87.57%), significantly outperforming the standard Hough transform and also surpassing LaneNet. Moreover, the detection efficiency is 26.3 m/s, which is 2.5 times faster than LaneNet and nearly 6 times faster than the standard Hough approach. This improvement is attributed to the combination of road‑area segmentation (which reduces search space), adaptive thresholds (which reduce false edges), and probabilistic Hough (which speeds up line voting). The integration of drone technology enables wide‑coverage imaging, allowing a single frame to contain over 100 meters of road length, thereby further boosting the effective throughput.
4.3 Qualitative Analysis
Visual inspection of the results shows that my method reliably extracts both solid and dashed lane markings, even in the presence of shadows, worn paint, and minor occlusions. The U‑Net segmentation effectively removes buildings, trees, and vehicles from the region of interest, preventing false detections. The color‑based extraction correctly distinguishes yellow from white markings. The adaptive Canny edge detector produces clean edges without over‑smoothing, and the probabilistic Hough transform correctly picks up dashed segments without merging them inappropriately. After slope filtering and fitting, the final lane lines align well with the ground truth. In contrast, the standard Hough transform often misses faint markings and generates many spurious lines, while LaneNet occasionally confuses road shoulders with lane markings due to limited context. The proposed method also shows robustness to varying illumination because of the white balance correction and the HLS color thresholding.
5. Discussion
Drone technology provides a unique vantage point that simplifies lane detection. Unlike vehicle‑based cameras that require complex perspective transformation to obtain a top‑down view, drone images are naturally orthorectified (assuming proper camera calibration and flight stabilisation). This eliminates projection errors and allows direct use of Euclidean geometry for line fitting. Additionally, the high revisit frequency of drone operations enables rapid map updates, which is crucial for dynamic environments such as construction zones or seasonal road changes. However, some challenges remain. For instance, dense tree canopy or bridge overpasses can cast strong shadows that degrade color segmentation. In such cases, additional preprocessing like histogram equalization or using near‑infrared bands could be beneficial. Another issue is the small size of lane markings relative to the image resolution; while centimeter‑level resolution is sufficient, lower‑altitude flights may incur higher operational costs. A trade‑off between altitude and ground sampling distance must be carefully managed. Despite these limitations, the experimental results confirm that the proposed pipeline is a practical and efficient solution for lane marking extraction in high‑definition mapping, fully embracing the advantages of drone technology.
6. Conclusion
In this paper, I have presented a complete framework for extracting lane markings from drone‑captured aerial images. The method integrates U‑Net road segmentation, HLS color space thresholding, adaptive Canny edge detection, and probabilistic Hough transform. Extensive experiments on real drone datasets demonstrate that the proposed approach achieves a recall of 86.23%, precision of 87.57%, and a detection efficiency of 26.3 m/s, outperforming both traditional Hough transform and the deep‑learning‑based LaneNet. The use of drone technology significantly reduces data acquisition costs while providing centimeter‑level accuracy and wide coverage. This work opens the door for frequent, low‑cost updates of high‑precision maps, which are essential for autonomous driving and smart city infrastructure. Future work will focus on integrating temporal information from video sequences to improve robustness against occlusions and weather conditions, and on extending the pipeline to detect other road elements such as arrows and stop lines. The synergy between drone technology and advanced image processing will continue to drive innovations in automated mapping and surveying.
