Ning Wang, Qu Jian · INTERNATIONAL SCIENTIFIC JOURNAL OF ENGINEERING AND TECHNOLOGY (ISJET) 2026 · 2026
DOI: 10.67765/isjet.v10i2.256045
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
This review paper examines explainable image captioning techniques for autonomous driving. It focuses on natural language explanations to improve transparency and explainability. We review more than 20 studies published from 2015 to 2024. These include foundational models, attention mechanisms, Transformer-based methods, and integrations with large language models. The review traces the progress from basic textual descriptions to chain-based reasoning and multimodal fusion. Because of resource limits and the need for low-cost experimentation, we chose the open-source DonkeyCar platform to replicate models. We replicated the ADAPT (Action-aware Driving Caption Transformer) model and compared it with prior methods using standard metrics: BLEU-4, CIDEr, and METEOR. The results show that ADAPT outperforms previous state-of-the-art methods, with CIDEr gains of 31.7 in narration and 33.1 in reasoning. We also successfully reproduced lane-following and obstacle-avoidance tasks on DonkeyCar. Multimodality supports more natural learning for future AI models, but accuracy is limited in most scenarios.
No comments yet — start the discussion below.