Visual Instruction Tuning for Drone Accident Forensics

Forensic Analysis Drone Forensics LLaVA Drone Accident.

Authors

  • Arda Surya Editya Department of Informatics, Institut Teknologi Sepuluh Nopember, Surabaya, East Java,, Indonesia
  • Tohari Ahmad
    tohari@its.ac.id
    Department of Informatics, Institut Teknologi Sepuluh Nopember, Surabaya, East Java,, Indonesia
  • Hudan Studiawan Department of Informatics, Institut Teknologi Sepuluh Nopember, Surabaya, East Java,, Indonesia

Downloads

The increasing use of drones in both commercial and personal use has led to a growing demand for effective forensic analysis following drone-related accidents. This research focuses on improving forensic analysis through the development of LLaVAFor, a fine-tuned version of the Large Language and Vision Assistant (LLaVA) model. The objective of this study is to enhance the interpretability of visual instruction tuning for drone accident forensics. LLaVAFor was developed by fine-tuning LLaVA via a specialized dataset of drone accident scenarios. The model's performance was evaluated via the BLEU score, a metric commonly used to assess machine translation and natural language processing models. The results demonstrated that LLaVAFor achieved superior BLEU scores compared with baseline models such as LLaVA, Google Gemini, and ChatGPT. It demonstrates its ability to provide more accurate and contextually relevant analyses. The key innovation in LLaVAFor is its ability to explain forensic findings in the context of drone accidents, making it a valuable tool for investigators. The results show that the model's fine-tuning process on drone-specific datasets enables it to offer detailed, domain-specific insights, improving the accuracy and reliability of forensic analyses in this field. Through these advancements, LLaVAFor represents a step forward in the integration of AI into drone accident investigations.

 

Doi: 10.28991/HIJ-2024-05-04-01

Full Text: PDF