Lane line detection method combining deep learning and image processing

Abstract

The invention designs a lane line automatic detection method based on combination of a YOLO algorithm and image processing. The method comprises an image acquisition module, a lane line identification module, an area screening module and a lane line marking module. Firstly, an image of the driving condition in front of a vehicle is collected through a vehicle-mounted camera; secondly, through a target detection algorithm and an optimal model weight obtained through training of a lane line training set, calibration of a rectangular area where a lane line is roughly located is obtained; thirdly, carrying out region screening by utilizing a mask; and finally, marking the lane lines through a traditional image processing mode. According to the method, a lane line detection method combining deep learning and a traditional image processing mode is established, the calculated amount is small, the accuracy rate is high, and the method is suitable for application of a carrier in many scenes.

Description

Technical Field

[0001] This invention relates to a lane detection method that combines deep learning and image processing, belonging to the field of target detection. Background Technology

[0002] With the rapid development of artificial intelligence and deep learning technologies, autonomous driving technology has gradually become a hot topic. The realization of autonomous driving technology relies heavily on accurate perception of the surrounding environment. Lane detection is a crucial component of autonomous driving systems, providing vehicles with critical environmental perception and location information. This is essential for achieving autonomous navigation and driving, and forms the foundation for automatic navigation and vehicle control. However, accurately detecting lane lines in complex road scenarios remains challenging. For example, traditional lane detection methods often rely on manually designed feature extraction methods, leading to incomplete representations of lane line features, a problem particularly pronounced in complex scenarios. Furthermore, these traditional image-based lane detection methods are often sensitive to changes in lighting, weather, and road conditions, exhibiting limited generalization ability across different environments. In addition, these traditional methods require extensive parameter adjustments and manual optimization in practical applications, resulting in higher labor and time costs. Deep learning-based lane detection methods not only require a large amount of training data but also suffer from computational complexity. Therefore, a combined approach integrating both methods can leverage their unique advantages. This paper proposes a lane detection method that combines deep learning and image processing. By changing the deep learning network structure and activation function of the activation layer, it combines various traditional image processing methods to solve some of the defects in current lane detection. Summary of the Invention

[0003] The technical problem solved by this invention is to address the shortcomings of existing lane detection methods in the field of autonomous driving by providing a combined method that improves lane detection accuracy, has a faster detection speed, and enhances the detection capability of the lane detection system.

[0004] The technical solution of this invention is: a lane line detection method combining deep learning and image processing, comprising the following steps: First, after acquiring images through an onboard camera, a deep learning lane line recognition system is built; second, the lane line anchor selection boxes identified by the built lane line recognition system are filtered for regions; finally, the lane lines are labeled using image processing methods such as color filtering and edge detection.

[0005] The first step is to build a deep learning-based lane line recognition model: After acquiring road images from an onboard camera, a deep convolutional neural network (CNN) model is built. This model learns image features through multiple convolutional and pooling layers, combined with appropriate activation functions and regularization techniques to enhance the ability to detect lane lines in complex road conditions.

[0006] The second step is to construct a region filtering system: Based on the output of the lane line recognition model, a region filtering system is designed to accurately divide and filter the detected lane line anchor boxes. Simultaneously, a mask is used to filter out the remaining areas of the image while preserving the original region, ensuring that subsequent image processing operations for lane lines are performed only in high-confidence regions, thereby reducing false detections.

[0007] The third step involves designing an image processing lane line labeler: Image processing methods such as color filtering and edge detection are used to further label the candidate lane line boxes after region filtering. These methods can effectively extract and identify the accurate location and shape of lane lines, providing crucial information for subsequent path planning and vehicle control.

[0008] The embodiments of the invention include, but are not limited to, the following steps:

[0009] In the first step, high-definition road images are first captured from an onboard camera as input data. Then, based on this data, a deep convolutional neural network (CNN) model is constructed. This model typically includes multiple convolutional layers, pooling layers, activation functions, and fully connected layers. Different sizes and numbers of convolutional kernels are used in the convolutional layers to extract features from the images, such as edges and textures. These features are crucial for lane line recognition. As the network depth increases, the convolutional layers can learn higher-level features, which are particularly important for lane line detection in complex road environments. Simultaneously, to enhance the model's generalization ability, the model of this invention includes an efficient spatial pyramid pooling layer. This pooling layer reduces overfitting by decreasing the spatial size and number of parameters of the data, while retaining important feature information. Replacing commonly used sigmoid and ReLU activation functions with rational fraction-type activation functions allows the model to learn complex nonlinear mapping relationships with almost no impact on prediction accuracy, significantly reducing computational complexity for subsequent calculations. Furthermore, to prevent overfitting during training, this invention also employs regularization techniques such as Dropout. Finally, we trained the model using a large dataset of manually labeled CULANE lane line images. By continuously adjusting the network structure and hyperparameters, we finally obtained a high-performance lane line recognition model.

[0010] In the second step, after the lane line recognition model outputs the detection results, the identified lane line anchor boxes need further processing. This is because the model may generate some false detections or false positives during the recognition process, which need to be eliminated by a region filtering system. First, a confidence-based filtering mechanism outputs a confidence score for each lane line anchor box, representing the probability that the box contains a lane line. We set a threshold; only anchor boxes with a confidence score higher than this threshold are retained. Next, we further filter using the spatial distribution characteristics and shape features of lane lines. By calculating parameters such as the angle and distance between adjacent anchor boxes, we determine whether they belong to the same group of lane lines. Only anchor boxes that meet these conditions are retained. This information helps us further eliminate false detections and false positives. After obtaining anchor boxes with a high probability of identifying lane lines, the region is preserved using a masking method.

[0011] In the third step, after region filtering, a set of high-confidence lane line candidate boxes can be obtained. Next, this invention needs to further annotate these candidate boxes to extract accurate lane line position and shape information. First, color filtering technology is used to filter the pixels in the candidate boxes. Since lane lines usually have specific colors, such as white and yellow, pixels belonging to lane lines can be filtered out by setting RGB color thresholds. This step helps us to more accurately locate the position of lane lines. Then, edge detection technology is used to extract the edge information of lane lines. Edge detection technology can detect areas in an image where brightness or color changes drastically, because these areas usually correspond to the edges or contours of objects, so this technology can be used to detect the edges of objects. For lane lines, their edge information is crucial for determining their accurate position and shape. Finally, the results of color filtering and edge detection are combined to annotate the lane lines. Specifically, this invention can annotate the pixels belonging to lane lines with specific colors or lines and calculate parameters such as the center line and width of the lane lines. This completes the automatic lane line detection task. Through the above three steps, we can achieve an efficient and accurate lane line detection method, providing strong support for the development of autonomous driving technology. Attached Figure Description Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation Implementation: This invention provides a lane detection method combining deep learning and image processing, comprising the following steps:

[0012] 1. Image Acquisition: Acquire road condition information in front of the vehicle using an onboard camera;

[0013] 2. Lane line recognition: Deep learning methods are used to anchor the location of lane lines;

[0014] 3. Region Filtering: Use a mask to filter the anchor selection box, retaining high-confidence regions;

[0015] 4. Lane line annotation: Lane lines are annotated using image processing methods such as color filtering, edge detection, and Hough transform.

[0016] This invention achieves efficient lane line detection by combining deep learning and image processing technologies, with high accuracy and adaptability.

Claims

1. A lane detection method combining deep learning and image processing, characterized in that, The lane line detection method includes: Image acquisition: capturing road condition information in front of a moving vehicle using an onboard camera; Lane line recognition uses deep learning methods to anchor the location of lane lines. Region filtering, used to filter out anchor selection boxes; Lane line marking is performed within the anchor selection box using methods such as color filtering, edge detection, and Hough transform.

2. The lane detection method combining deep learning and image processing according to claim 1, characterized in that, The lane line recognition method includes: using a network structure different from the original YOLOv5 object detection; adding an attention mechanism to the network structure; and replacing the activation functions in the activation layer by replacing some sigmoid and ReLU functions with rational fractions, which greatly reduces the computational complexity for subsequent calculations without affecting the prediction accuracy.

3. The lane detection method combining deep learning and image processing according to claim 1, characterized in that, The region filtering method involves using a different approach than manually setting filtering regions. Instead, it combines rectangular anchor selection boxes identified by deep learning as a reference for region filtering, and then filters regions based on the geometric features of lane lines within these anchor selection boxes.

4. The lane detection method combining deep learning and image processing according to claim 1, characterized in that, The lane line annotation process includes: based on prior knowledge that lane lines are often presented in white or yellow, this step first sets an appropriate threshold in the RGB color space for preliminary screening based on this feature, then converts it to a grayscale image and uses a Gaussian filter for noise reduction, and finally uses the Canny operator for edge detection and combines it with Hough transform to complete the lane line annotation.

Images