Using an Attention Mechanism for Rapid Lane Detection

Abstract

Lane recognition, a crucial subtask in autonomous driving, has recently shifted from using traditional image processing to a neural network technique based on deep learning. On the other hand, early deep learning approaches failed to match real-time needs due to their reliance on pixel-level semantic segmentation and therefore massive network architectures. A novel network architecture denoted by UFAST and built on preset rows is suggested as a solution to the real-time challenge. This network model's design achieves real-time performance by drastically reducing the network parameters. We include an attention mechanism into the model based on observations of actual human driving behaviors in order to enhance the lane recognition performance within this framework. Finally, we enhance the model framework's performance in less-than-ideal conditions by nearly 1.9% and increase the number of parameters by less than 0.2% of the UFAST. This is done by artificially scheduling the input image data to the lower part of the view, where the lanes are typically located. But actual road data is complicated, and using the same method for all the view data would result in duplicate or missing data. While the ResNet-18 [6] architecture achieved a classification accuracy of 95.87% on the TuSimple dataset [5], the CULane dataset [2] achieved just 68.4% accuracy in its trials, which is somewhat inadequate. This is because, under less-than-ideal circumstances, not only do road lane details likely to be missing from the CULane datasets because to the datasets' inherent complexity and variability, but the system framework's architecture also contributes to this problem.