High-Resolution Fusion Lane Detection Algorithm Based on Model Ensemble

Lane line detection is still a challenging task due to the challenges coming from the complexity of drone aerial images such as complex lane lines,fine-grained feature,class imbalance,etc.Therefore,a lane line detection algorithm based on high-resolution fusion convolution network is proposed.Firstly,the convolution module and up sampling module of full convolution network are improved by using high-resolution fusion structure and bilinear interpolation algorithm.Then,according to the idea of model ensembling,the improved model architecture is used as the foreground-background semantic segmentation model and the multi-category semantic segmentation model,which is used to solve the problem of lane line detection step by step,and the two models are trained by the joint loss function composed of threshold cross entropy loss and Lovasz loss.Finally,the locally region-growth algorithm is used to supplement the details of the detected results.The experimental results show the algorithm achieves 0.5484 mean intersection over union and 0.9931 pixel accuracy in the customized drone aerial dataset of 15 types of lane lines,and the prediction speed of 512×512 resolution image on NVIDIA Tesla V100 reaches 23.08 frame per second.