Lane Detection Based On Geometric Regularization Contraint

Lane detection,as an important part in intelligent transportation,plays an important role in Advanced Driving Assistant System(ADAS)and Automatic Violation Detection.Lane detection needs to extract key features based on input environment data and detect lane areas with a high frame rate and accuracy.There are some mature lane detection algorithms in ideal situation.However,due to the diversity of scenes and image quality,existing algorithms do not guarantee lane detection accuracy in all scenarios.Lane detection is one of the most challenging problems.Currently,multi-sensors algorithms can improve the accuracy of lane detection,but it also increases equipment cost and limits the practicality of algorithms.On the other hand,some computer vision algorithms and deep learning algorithms consider texture and structure information from lanes but ignore the edge features provided by lane boundaries.In this paper,we mainly discuss lane detection on a single frame.We design a multi-task neural network to conduct lane area detection and lane boundary detection at the same time,which makes full use of texture and boundary features from original images.The features from both tasks complement each other to improve accuracy.We use semantic segmentation network based on convolution-deconvolution to detect lane areas in a single frame.Convolution layers extract image features,while deconvolution layers detect lane areas in a pixel-wise manner.More important,we introduce a novel loss function to impose geometric constraint on the lane area and the lane boundary.It improves the performance of our network structure.With regard to the above limitations of current methods,this paper proposes two improvements to enhance features between two sub-networks: inter-link structures and structure loss functions.The first improvement is to build a multi-task network and detect lane areas and lane boundaries in an end-to-end manner,thereby ensuring that the convolution layers can extract good feature information.To obtain feature fusion between two tasks,we introduce an inter-link network structure between sub-networks.Based on boundary information extracted by lane boundary detection,lane area could be detected with higher accuracy.On the other hand,area information from lane detection can guide lane boundary detection.Experiments show that our network structure outperforms other methods,including the state-of-art single and multi-task lane detection networks.The second improvement is to build structure loss functions and constrain the training process with geometric relationship.Current loss functions calculate pixel-wise deviation of results from ground truth and ignore overall geometric structure distortions.In contrast,we measure boundary consistency by comparing the extracted boundaries from lane area detection results and the corresponding lane boundary ground truth.Meanwhile,we evaluate area consistency from lane boundaries and compare it with the ground truth to measure lane area consistency.Our single frame lane detection algorithm makes full use of prior geometric relationship between lanes and lane boundaries.We promote feature fusion to generate lane detection results in an end-to-end manner.Our network structure outperforms the state-of-arts in several metrics(accuracy,recall and so on).Meanwhile,experiments also show that our algorithm is more robust and can detect lane areas in several different scenes with promising performance.