Development and testing of an image transformer for explainable autonomous driving systems

Purpose Perception has been identified as the main cause underlying most autonomous vehicle related accidents. As key technology in perception, deep learning (DL) based computer vision models are generally considered to be black boxes due poor interpretability. These have exacerbated user distrust and further forestalled their widespread deployment practical usage. This paper aims develop explainable DL for driving by jointly predicting potential actions with corresponding explanations. The can not only boost trust autonomy but also serve a diagnostic approach identify any model deficiencies or limitations during system development phase. Design/methodology/approach proposes an end-to-end on “Transformer,” state-of-the-art self-attention (SA) model. maps visual features from images collected onboard cameras guide explanations, achieve soft attention over image’s global features. Findings results demonstrate efficacy of proposed it exhibits superior performance (in terms correct prediction explanations) compared benchmark significant margin much lower computational cost public data set (BDD-OIA). From ablation studies, SA module outperforms other mechanisms feature fusion generate meaningful representations downstream prediction. Originality/value In contexts situational awareness driver assistance, perform alarm both human-driven vehicles because is capable quickly understanding/characterizing environment identifying infeasible actions. addition, extra explanation head provides channel sanity checks guarantee that learns ideal causal relationships. provision critical systems.