Research On Real-time Lane Detection Algorithm Based On Deep Learning

In recent years,my country’s economic development has been changing with each passing day,which has prompted the car ownership to show a trend of steady growth every year.Various traffic accidents on highways occur from time to time,causing a enormous threat to our life and health.According to reports,deviating from the driving lane during driving is a major factor leading to traffic accidents.Therefore,relevant scientific researchers regard the automatic positioning of the lane as a major task in the field of automatic driving,and the key is the lane detection.Lane detection is a fundamental problem for applying computer vision to autonomous or assisted driving.Early lane detection tasks were mainly achieved by traditional image processing methods.However,the reality is that lines are always in a complex road scene such as severe occlusion,unclear lanes and poor lighting conditions.This brings huge difficulties to traditional image processing methods.In comparison,deep segmentation methods have stronger semantic representation capabilities and can perform higher-level semantic analysis on lanes,thus gradually becoming the mainstream.Unfortunately,although this convolutional neural network-based method has superior scene understanding ability,it is still difficult to deal with lane lines with long structural regions and even may be occluded.In addition,lane line detection,as the basic module in the intelligent vehicle environmental perception system,has less resources allocated,so it is need to reduce the cost of the detection algorithm as possible.This paper proposes a deep learning-based real-time Lane detection method,which is able to achieve faster speed and higher performance compared to most current state-of-the-art methods.This paper proposes a lane detection model based on row classification and attention mechanism,Lane-SA.Specifically,this paper adopts a linebased classification method,which aggregates global information by fusing an attention mechanism of channel attention in series with spatial attention,and aggregates global features and local features in the detection head,and combines semantic segmentation as an auxiliary branch.In this paper,extensive experiments are performed on two public datasets respectively,and compared with the current state-of-the-art methods,the efficiency trade-off of detection is discussed.Finally,this paper carries out the postprocessing work of adaptive fitting on the feature points of the lane detected by the model.In conclusion,the main contributions are:（1）.A simple,fast and effective lane detection method is proposed,which can outperform state-of-the-art methods on both large and complex datasets.This method adopts a method based on row classification,and redefines the task of lane line detection as the selection and classification of lane positions based on the direction of the row,with a simple and lightweight structure,thus ensuring the real-time detection.In order to better solve the problem of no visual cues and further improve the detection efficiency of the model,this paper combines spatial domain and channel domain attention mechanisms,fuses a lightweight attention mechanism to aggregate global features,and fuses global features and local features,which also have potential utilization value for related detection objects in other fields.（2）.A sub-regional adaptive lane line fitting method based on feature point attention is proposed,which can well fit the detected lane line feature points into lines.In this paper,the lane scene image is divided into three different visual fields,and the line and curve fitting algorithms involving the attention of feature points are used to process the lane lines of different visual fields,and finally the parametric equation of the fitting result is output.Compared with the single straight line fitting or curve fitting method,the adaptive fitting method can better reflect the characteristics of the lane scene.Combined with the feature point attention,it improves the anti-noise ability of the feature points,so it can also perform better.The algorithm in this paper comprehensively considers the accuracy and speed of detection,and has achieved fairly good results in these two key indicators,which shows that the algorithm in this paper has certain feasibility and practicability.