Lane detection algorithm based on multi-head self-attention and multi-level feature fusion

Abstract

Lane detection is a crucial environmental sensing technique that is used in advanced driving assistance systems and automatic driving. The research on this issue has significant practical value. Aiming the current lane detection algorithm could not solve the problems of the local receptive field and detail feature loss, we introduced the multi-head self-attention module in Transformer into the encoder and decoder to obtain the global receptive field while solving the problem of detail feature loss with the multi-level feature fusion decoder. The proposed algorithm has been compared with the ERFNet model in the CULane dataset, and the detection accuracy has improved by 3.9 percentage points. The detection accuracy in the Tusimple dataset is 96.51%. Introducing a multi-head self-attention module increases the feature selection effect of the attention mechanism in the coding and decoding process. It provides a new solution for the lane detection algorithm.