A Lane-Changing Detection Model Using Span-based Transformer

Lane-changing is an important driving behavior and unreasonable lane changes can potentially result in traffic accidents. Therefore, it is urgent to develop lane change detection systems, especially which could work in the initial phase of lane change. In this paper, a novel Transformer-based model, Lane-changing Span-based Transformer (LS-Transformer) is proposed for lane-changing behavior detection from front view videos. Firstly, with Span-based Dynamic Convolution (SDConv), LS-Transformer is capable of integrating the convolution and self-attention to efficiently capture both global and local dependencies with decreased redundancy. Secondly, the lane boundary based distance features and visual scene features can be generated by the LS-Transformer simultaneously. Finally, a Long Short-Term Memory (LSTM) neural network which models temporal dependencies is designed to learn the co-occurrence features and detect lane-changing behaviors. The model is evaluated on a challenging self-collected data set from real driving scenarios, and the experimental results reveal that the proposed LS-Transformer outperforms the other advanced models with faster speed.