🚘 DriveSOTIF: Advancing Perception SOTIF Through Multi-Modal Large Language Models

---

👩‍💻 Authors

Shucheng Huang, Freda Shi, Chen Sun, Jiaming Zhong, Minghao Ning, Yufeng Yang, Yukun Lu, Hong Wang, Amir Khajepour

---

🧠 Overview

DriveSOTIF is the first work to address perception-related Safety of the Intended Functionality (SOTIF) challenges in autonomous driving using Multi-Modal Large Language Models (MLLMs).
We introduce the DriveSOTIF dataset, benchmark state-of-the-art MLLMs (Qwen2.5-VL, InternVL3), and provide:

• 🔹 Sample code for dataset generation
• 🔹 Ensemble YOLO-based perception enhancement
• 🔹 Uncertainty estimation modules for perception confidence

---

🧾 Abstract

Human drivers possess spatial and causal intelligence, enabling them to perceive driving scenarios, anticipate hazards, and react to dynamic environments.
In contrast, autonomous vehicles lack these abilities, making it challenging to manage perception-related SOTIF risks, especially under complex or unpredictable driving conditions.

To address this gap, we fine-tune multimodal large language models (MLLMs) on a customized dataset designed for perception-related SOTIF scenarios. Benchmarking results demonstrate:

• +11.8% improvement in close-ended VQA accuracy
• +12.0% increase in open-ended VQA scores
• 0.59 s average inference time per image (real-time capable)

Real-world case studies from Canada and China validate that fine-tuned models can identify subtle safety risks that even experienced human drivers may overlook.

This work represents the first domain-specific MLLM fine-tuning for the SOTIF domain in autonomous driving.

---

Dataset and code release:

Since DriveSOTIF is built upon the PeSOTIF dataset, we wont be able to release the dataset directly due to licensing issues. However, we release sample codes that you can refer to build your own dataset. You can find the released code in the code folder. The code folder contains the following parts:

Dataset generation scripts (code/gen_dataset.py):

This is demo code to generate a dataset with caption, question and answer pairs for fine-tuning MLLMs. It contains the following features:

1. Image caption generation, evaluation and metadata extraction
2. Question generation, evaluation and metadata extraction
3. Answer generation, evaluation and metadata extraction
4. Overall evaluation and post-processing

For the evaluation part, we currently check every single caption, question and answer and regenerate failed entries with GPT-4o/GPT-5 to ensure the quality of the generated dataset. You can sample (eg. 2-3 out of 5) generated captions, questions and answers and evaluate them with GPT-4o/GPT-5 to save API credits.

Please be aware that this released demo code is built using GPT and Claude code for code streamlining purposes, and thus the code may look less human like (eg. ~1500 lines of code in a single file). However, the core logic and structure of the code is still human written. The code is modularized and well commented for easy understanding. Also, please note that the code is only a demo version for dataset generation and don't include the switching LLM backends. But it is quite straightforward to integrate with any LLM backend of your choice. Note that using proprietary LLMs such as GPT-4o/GPT-5 etc might consume a significant amount of API credits, especially when processing a large number of images. We recommend testing the code with a small batch of images first to estimate the cost before scaling up.

Ensemble Yolo models for perception (code/sotif_emsemble.py):

We provide sample codes to ensemble multiple (5) yolo models to enhance perception performance and estimate uncertainty based on the SOTIF entropy method proposed in "SOTIF entropy: Online SOTIF risk quantification and mitigation for autonomous driving". This is not an official implementation of the SOTIF entropy method, but our own implementation based on the method described in the paper.

Please be aware that in the original SOTIF entropy paper, the authors used YOLOv5 models, which officially support ensembling. However, in this released code, we used YOLOv8 models, which do not provide complete category-level confidence outputs. To obtain the category-level confidence outputs, the hard way is to modify the YOLOv8 source code and expose the required outputs. We took a shortcut by using the bounding box-level confidence outputs and aggregating them to get category-level confidence outputs. This is essentially an conservative approximation of the original SOTIF entropy method, which may lead to slightly different uncertainty estimates. However, we found that this approximation still works well in practice and provides a good balance between performance and complexity.

For more details, please refer to the code provided and our supplementary material. Below is a brief summary of the method and our implementation:

1. Aggregate ensemble confidence.
   For each object cluster (detections across models that overlap and share class), compute the mean confidence
   s_bar = mean([s_m for each model m]).

2. Base (binary) entropy.
   Treat the problem as “detected class” vs “all others” with probability p = s_bar.
   Binary entropy:
   H_b(p) = -( p * log p + (1 - p) * log(1 - p) ) .

3. Approximate other classes (sampling heuristic).
   Since YOLOv8 hides the rest of the class scores, we:

   • Sample several plausible class-probability vectors that distribute 1 - s_bar across the remaining K-1 classes (using a helper like random_prob(s_bar, K)).
   • Compute the binary entropy for each sampled top-class probability and take the maximum (worst case).
   • Combine with the base entropy to get a conservative SOTIF entropy:
     H_SOTIF = max_n H_b(p^(n)) + H_b(s_bar).

4. Weak consensus adjustment.
   If fewer than 5 models support a cluster, inflate uncertainty:
   H_final = H_SOTIF * (1 + 0.1 * (5 - c)), where c is the number of contributing detections. 5 can be tuned based on the number of models in the ensemble.

5. Uncertainty buckets.

   • Low: H < 1.2
   • Medium: 1.2 <= H < 1.6
   • High: H >= 1.6

Citation

If you find DriveSOTIF useful in your research, please cite:

@article{huang2025drivesotif,
  title={DriveSOTIF: Advancing SOTIF Through Multimodal Large Language Models},
  author={Huang, Shucheng and Shi, Freda and Sun, Chen and Zhong, Jiaming and Ning, Minghao and Yang, Yufeng and Lu, Yukun and Wang, Hong and Khajepour, Amir},
  journal={IEEE Transactions on Vehicular Technology},
  year={2025},
  publisher={IEEE}
}

Acknowledgements

We would like to acknowledge the financial support of the Natural Sciences and Engineering Research Council of Canada (NSERC) and Vector Institute. We also acknowledge the research grant provided by OpenAI and computing resources provided by the Vector Institute.

License:

The PeSOTIF and CADC dataset are accessed under CC BY-NC-SA 4.0 license. Proprietary LLMs, such as GPT4, GPT4o, etc. are accessed under their licenses, terms and conditions. Open-source LLM models and LVLM, such as Blip, Blip2, LLAVA, Qwen2-VL, etc., are accessed under their corresponding licenses.

Copyright notice for the journal paper:

©2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.