VisionGPT

KO | EN

gitlite — search

#hacktoberfest #react #ai #typescript #android #machine-learning #rust #linux #docker #macos #awesome #pytorch #security #open-source

(★ 40)

LLM-Assisted Real-Time Anomaly Detection for Safe Visual Navigation

VisionGPT Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# VisionGPT: LLM-Assisted Real-Time Anomaly Detection for Safe Visual Navigation

## Use
Copy/Download the repository, replace the API key in OPENAI_API_KEY.yaml, and run the demo.py.

## Demonstration
<div align="center">
    <img src="./Figures/JP_1.gif" alt="Framework" style="width: 60%;">
</div>

See our another project for the movement prediction: <a href="https://github.com/JiayouQin/H-Splitter" target="_blank">H-Splitter</a>

## Overview
This project explores the potential of Large Language Models(LLMs) in zero-shot anomaly detection for safe visual navigation.

<br>

With the assistance of state-of-the-art real-time open-world
object detection model Yolo-World and specialized prompts, the proposed framework can identify anomalies within camera-captured frames that include any possible obstacles, then generate
concise, audio-delivered descriptions emphasizing abnormalities, assist in safe visual navigation
in complex circumstances.

Moreover, our proposed framework leverages the advantages of LLMs
and the open-vocabulary object detection model to achieve the dynamic scenario switch, which
allows users to transition smoothly from scene to scene, which addresses the limitation of traditional
visual navigation.

Furthermore, this project explored the performance contribution of different prompt
components provided the vision for future improvement in visual accessibility and paved the way
for LLMs in video anomaly detection and vision-language understanding.

## Method

### Yolo-World
We apply the latest Yolo-world for the open-world object detection task to adapt the system in any scenario any situation. The detection classes are generated by GPT-4 and can be replaced dynamically.

### GPT-3.5
We apply GPT-3.5 for fast response and low cost. We have tested GPT-4 and GPT-4V but found them not financial-friendly.

### H-splitter
We implemented an H-splitter to assist object detection and categorize the objects into 3 different types based on the priority.

See our another project for more info: <a href="https://github.com/JiayouQin/H-Splitter" target="_blank">H-Splitter</a>

## Experiments
We use Yolo-World with the H-splitter for universal object detection. For any object that falls (a)in Area 3 or (b)in Area 1/2 with 15% of window size, we record the corresponding frame as anomalies. We set this Yolo-World-H setting as the ground truth for the benchmark.

### System Sensitivity
We pre-set the system with 3 different sensitivities to report the emergency: low, normal, and high. We find that the low system sensitivity is good for daily use due to the low false alarm rate.

### Detection accuracy
We compare the VisionGPT with low system sensitivity with the ground truth to evaluate its performance. We find that VisionGPT has high Accuracy and prefers less False Positive (unnecessary reports).

### System Performance & Compatibility
<div align="center">
    <img src="./Figures/fps.png" alt="System performance" style="width: 80%;">
</div>

### Cost Evaluation
<div align="center">
    <img src="./Figures/cost.png" alt="Financial cost test" style="width: 70%;">
</div>

### Acknowledgements:

Please cite our work if you find this project helpful.
```bibtex
@article{wang2024visiongpt,
  title={VisionGPT: LLM-Assisted Real-Time Anomaly Detection for Safe Visual Navigation},
  author={Wang, Hao and Qin, Jiayou and Bastola, Ashish and Chen, Xiwen and Suchanek, John and Gong, Zihao and Razi, Abolfazl},
  journal={arXiv preprint arXiv:2403.12415},
  year={2024}
}
```

VisionGPT

VisionGPT Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

VisionGPT

VisionGPT Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README