Random Structural Data Augmentation for Lane Detection

Lane detection is one of the fundamental yet important tasks in autonomous driving, which provides further clues for drivable regions detection and lane departure decision. Nowadays, Convolutional Neural Network(CNN) based lane detection methods have achieved a great success due to its strong contextual representation learning ability. Nevertheless, these methods still suffer dramatic performance degradation under challenging driving scenarios where lane occlusion and various extreme light conditions may occur. In fact, the performance of a lane detection model is closely related to the quality of the data representation. Normally, lanes are spatially continuous and appear on the ground. Therefore, we believe proper utilization of these structure prior in the data augmentation should lead to better detection precision under challenging scenarios. To this end, this paper explores the effect of a series of random structural data augmentation methods when applied to a row anchor based lane detection network. The experimental results confirm that structural data augmentations like ’Extending’ and ’Cutout’ can help network focus on the structural clues and improve lane detection performance by a large margin especially in challenging scenarios lacking visual clues. We also elaborate intuitions behind these methods and they can be easily applied to many other lane detection algorithms without effort. The achieved results have already been implemented in Kaizhou District, Chongqing.