CN-121999449-A - Lane line detection method combining deep learning and image processing

CN121999449ACN 121999449 ACN121999449 ACN 121999449ACN-121999449-A

Abstract

The invention designs a lane line automatic detection method based on the combination of a YOLO algorithm and image processing, which comprises an image acquisition, a lane line identification module, a region screening and a lane line marking module, wherein an image of the running condition in front of a vehicle is acquired by utilizing a vehicle-mounted camera, the calibration of a rectangular region where a lane line is approximately located is obtained through a target detection algorithm and the optimal model weight obtained through training of a lane line training set, the region screening is performed by utilizing a mask, and finally the lane line marking is performed through a traditional image processing mode.

Inventors

CHEN HAO
ZHANG HONG
LI YAWEI
LI XULIANG
YANG YIFAN
ZHANG JIAWEI

Assignees

苏州智臻威视光电科技有限公司

Dates

Publication Date: 20260508
Application Date: 20240627

Claims (4)

1. The lane line detection method combining deep learning and image processing is characterized by comprising the following steps of: Image acquisition, namely acquiring road condition information in front of a running vehicle through a vehicle-mounted camera; lane line identification, namely anchoring the existence position of the lane line by a deep learning method; Area screening, which is used for screening out an anchor selection frame; And marking lane lines, namely marking the lane lines in the anchor frame by means of color screening, edge detection, hough transformation and the like.
2. The lane line detection method combining deep learning and image processing according to claim 1, wherein the lane line recognition includes using a network structure different from the original YOLOv target detection, adding an attention mechanism into the network structure, replacing an activation function in an activation layer, and replacing part of sigmoid functions and Relu functions in the activation function with a rational component type, so that the computational complexity is greatly reduced to facilitate the calculation of a subsequent process under the condition of hardly affecting the prediction accuracy.
3. The method for detecting lane lines by combining deep learning and image processing according to claim 1, wherein the region screening comprises the steps of using a method different from a common manual setting screening region, and carrying out region screening by taking a rectangular anchor frame identified by the deep learning as a reference, and carrying out region screening according to geometric characteristics of the lane lines on the basis of the anchor frame.
4. The method for detecting the lane line by combining the deep learning and the image processing according to claim 1 is characterized in that the lane line labeling comprises the steps of setting corresponding thresholds in an RGB color space according to the characteristics to perform preliminary screening through priori knowledge that the lane line is always white or yellow, converting the thresholds into gray images, performing denoising processing by adopting a Gaussian filter, and finally selecting a Canny operator to perform edge detection and combining Hough transformation to realize the labeling work of the lane line.

Description

Lane line detection method combining deep learning and image processing Technical Field The invention relates to a lane line detection method combining deep learning and image processing, and belongs to the field of target detection. Background With the rapid development of artificial intelligence and deep learning technologies, automatic driving technology is becoming a popular topic today. The realization of the automatic driving technology is not separated from the accurate perception of the surrounding environment, the lane line detection is a vital ring in an automatic driving system, provides key environment perception and position information for the vehicle, is vital for realizing the autonomous navigation and the running of the vehicle, and is the basis for realizing the automatic navigation and the vehicle control. However, achieving accurate detection of lane lines in complex road scenarios still has not little difficulty. For example, conventional lane line detection methods typically rely on feature extraction methods that require manual design, which can lead to incomplete representation of lane line features, and this problem is particularly pronounced in complex scenarios. Meanwhile, the traditional detection methods, namely the lane line detection method based on the image characteristics, are sensitive to changes of conditions such as illumination, weather, roads and the like, and have limited generalization capability in different environments. In addition, the traditional detection method requires a large amount of parameter adjustment and manual optimization processes in practical application, so that more labor cost and time cost are required in practical application, and the deep learning lane line detection method generally requires a large amount of training sample data in calculation and has the problem of complex calculation. Thus, a combined lane line detection method combining both methods can exert unique advantages of both. The lane line detection method combining the deep learning and the image processing combines a plurality of traditional image processing methods to solve some defects existing in the current lane line detection while changing the network structure of the deep learning and the activation function of the activation layer. Disclosure of Invention The invention solves the technical problem of providing a combination method aiming at the defect of the existing lane line detection method in the automatic driving field, improves the detection precision of the lane line, has higher detection speed and enhances the detection capability of a lane line detection system. The technical scheme is that the lane line detection method combining the deep learning and the image processing comprises the following steps of firstly, building a lane line recognition system of the deep learning after an image is acquired through a vehicle-mounted camera, secondly, carrying out area screening on a lane line anchor frame recognized by the built lane line recognition system, and finally, realizing the marking work of the lane line through the image processing methods such as color screening, edge detection and the like. Firstly, a deeply-learned lane line recognition model is built, namely a deep Convolutional Neural Network (CNN) model is built after a road image is acquired from a vehicle-mounted camera. The model learns image characteristics through a plurality of rolling layers and pooling layers, and combines proper activation functions and regularization technologies to enhance the detection capability of lane lines under the condition of complex roads. And secondly, constructing a region screening system, namely designing the region screening system on the basis of the output result of the lane line identification model, and carrying out accurate region division and screening on the detected lane line anchor frame. Meanwhile, the mask is used for filtering the residual region of the image on the premise of keeping the region, so that the follow-up image processing operation for the lane line is only performed in a region with high confidence coefficient, and false detection is reduced. And thirdly, designing an image processing lane line marker, namely performing further marking work on the lane line candidate frames subjected to the region screening through image processing methods such as color screening, edge detection and the like. The method can effectively extract and identify the accurate position and shape of the lane line and provide key information for subsequent path planning and vehicle control Embodiments of the invention include, but are not limited to, the following steps: In the first step, first, a high-definition road image is captured as input data from an in-vehicle camera. Then, based on these data, the present invention constructs a deep Convolutional Neural Network (CNN) model. This model typically contains multiple convolution layers, pooling layers, ac