A deeplabv3+-based Investigation on Lane Line Detection

This study offers a deeplabv3+ based lane line identification approach that splits the lane line detection challenge into semantic segmentation and instance segmentation tasks in order to overcome the vehicle occlusion problem in lane line detection. The Atrous Spatial Pyramid Pooling structure can fully utilize the information from different sensory fields, enhance the network's extraction of contextual information, and strengthen the network's reasoning ability to solve the problem of lane lane not being fully detected when the vehicle is obscured. The overall Xception-based deeplabv3+ network, with its depth-separable convolution method, can optimize the network computation speed. A combined loss function approach utilizing the Focal-loss function and Tversky-Ioss function is presented to address the issue of imbalanced positive and negative data in lane line detection. The Tusample dataset is used for experimental validation. The experimental findings demonstrate that the deeplabv3+-based lane line identification approach has an accuracy of 0.9711, providing a superior solution to the challenge of lane line recognition in the presence of obstructed cars.