Residual Network Based Curve Enhanced Lane Detection Method

With the rapid economic development,cars benefited the public and also caused a large number of traffic accidents.Advanced Driver Assistance System(ADAS)and unmanned driving technologies for automobiles have great potential in assisting driving and improving driving safety,so they have received extensive concern.As a basic and important part of ADAS and automatic unmanned driving,lane detection has become a research hotspot,but further research is still needed in the areas of curve detection and " no-visual-clue" lane detection.The traditional lane detection methods usually first perform a series of image preprocessing,and then perform fitting based on the extracted lane image features,and draw the line at the corresponding position of the original image when outputting the detection result.Although the traditional image processing algorithm detects lane lines in a simple and fast way,it is greatly affected by external factors,such as illumination,adjacent vehicles,and indicator arrows in the middle of the lane.Lane detection based on deep learning can learn more lane features by constructing multi-hidden-layer learning model and training massive data,so as to improving the accuracy of detection.As the application of deep learning in many fields has achieved breakthroughs,more and more researchers regard the lane line detection problem as pixel-level segmentation problems,and use deep models such as Convolutional Neural Networks to solve.However,none of the existing methods pay special attention to the detection of curves,resulting in poor detection effect of curve scene;secondly,the current deep models’ structures are relatively large and require more computing resources,which is difficult to meet the needs of real-time lane detection.Aiming at the current problems that existed in the deep learning methods.Based on the improvement of the curve detection effect,and comprehensively considers the detection speed,this thesis proposes Residual Network Based Curve Enhanced Lane Detection Method.This method uses the main framework of the Residual Network and realizes the curve enhancement by adding the curve structure constraints to the loss function.Since straight lines can be regarded as a special case of curves,the method proposed in this paper is also suitable for straight lanes and has good universality.On the other hand,in order to reduce the complexity of the model,the weights pruning technique is used to reduce the model.In order to verify the effectiveness of the method,experiments were performed on the conventional lane detection dataset Tusimple and the self-built curve test set.The results show that the model with the curve structure constraint in the loss function has improvements over the existing model in both the Tusimple dataset and the curve dataset in terms of accuracy,recall,and the F1 metric.This shows that the curve enhancement strategy proposed in this paper can not only effectively improve the algorithm performance in the curve scene,but also does not affect the detection performance of straight lanes.After adding the pruning strategy,the algorithm greatly reduces the calculation time without significant performance degradation,which is more in line with actual production requirements.This paper adopts the mainframe of the Residual Network of UFLD model,realizes curve enhancement by adding curve structure constraints to the loss function,and uses weights pruning technology to reduce the model,proposes a curve enhancement model for efficient lane detection.This model obviously improves the performance of the algorithm in the curve scene,greatly reduces the calculation time,has a good application prospect.