Research On Lane Detection Algorithm Based On Deep Learning

With the acceleration of urbanization and the increase of car ownership,the urban traffic burden is increasing.In order to avoid road accidents and ensure the safety of driving,vehicle driving assistance system has become a new research hotspot.Lane line detection can effectively guide the vehicle to drive in the correct area.As the core of vehicle driving assistance system,the speed and accuracy of lane line detection are very important.The traditional algorithm of lane detection needs to adjust the parameters manually,which has heavy workload,poor generalization ability and poor performance in the face of complex traffic environment.Lane detection algorithm based on deep learning has strong representativeness and learning ability,and has higher speed and accuracy than traditional algorithms.Segmentation-based algorithms have become the mainstream in deep learning algorithms,but segmentation algorithms still have some limitations.Firstly,the segmentation based algorithm needs to extract the features of the whole image pixel by pixel,but the lane line occupies only a small part of the pixels in the whole image.The segmentation method will greatly increase the amount of calculation,which will lead to the slow speed of detection,and the receptive field of the segmentation based method is limited and the global features are not used.Therefore,UFSA(Ultra Fast Structure-aware)algorithm came into being.Aiming at the problem of large amount of calculation in the segmentation algorithm,the algorithm based on UFSA proposes line selection to extract lane line features,which greatly reduces the amount of calculation of the model and improves the detection speed.A new lane loss is also proposed to model the position relationship of lane points,which solves the problem of no visual clues in lane line detection.Therefore,the algorithm is widely used in the field of lane line detection.While UFSA algorithm in network convolution and pooling will lose important information boundary information is not sensitive enough and so on this paper proposes a lane detection algorithm based on improved UFSA network structure and loss function.FCASPP(Frequency Channel Attention Spatial Pyramid Pooling)is defined as FCASPP(Frequency Channel Attention Spatial Pyramid Pooling),which can extract more useful and detailed features to suppress noise information in large receptive field.This paper also introduces L-Dice(Lane Dice Loss)function into the loss function,which pays more attention to the information of lane boundary than Softmax function,such as color,texture and illumination.Through training and experiments on Tu Simple data set and CULane data set,the experimental results show that the lane line detection algorithm based on improved UFSA network structure and loss function proposed in this paper improves the detection accuracy by 0.21% and 1.7% respectively compared with the original model,and the speed reaches 298.4fps.The new algorithm proposed in this paper pays more attention to the characteristics of curve loss than the original curve detection model,which shows that the new algorithm proposed in this paper is more effective than the original curve detection model.