Efficient spatial and channel net for lane marker detection based on self-attention and row anchor

Abstract

Lane detection is an important component of advanced driving aided system (ADAS). It is a combined component of the planning and control algorithms. Therefore, it has high standards for the detection accuracy and speed. Recently several researchers have worked extensively on this topic. An increasing number of researchers have been interested in self-attention-based lane detection. In difficult situations such as shadows, bright lights, and nights extracting global information is effective. Regardless of channel or spatial attention, it cannot independently extract all global information until a complicated model is used. Furthermore, it affects the run-time. However trading in this contradiction is challenging. In this study, a new lane identification model that combines channel and spatial self-attention was developed. Conv1d and Conv2d were introduced to extract the global information. The model is lightweight and efficient avoiding difficult model calculations and massive matrices, In particular obstacles can be overcome under certain difficult conditions. We used the Tusimple and CULane datasets as verification standards. The accuracy of the Tusimple benchmark was the highest at 95.49%. In the CULane dataset, the proposed model achieved 75.32% in F1, which is the highest result, particularly in difficult scenarios. For the Tusimple and CULane datasets, the proposed model achieved the best performance in terms of accuracy and speed.