Enhancing Road Safety with AI-Powered Accident Detection
Objective
The objective of this project is to develop an AI-driven system that detects accident scenes from images captured by CCTV footage. By leveraging advanced machine learning techniques, we aim to improve response times to road incidents, thereby enhancing overall road safety.
Data Sample
We utilized the
Accident Detection from CCTV Footage dataset from Kaggle. This dataset contains annotated images from CCTV footage, showcasing various accident scenarios.
Sample Data
Here’s a sample from the dataset:
| Image | Label |
|---|
| ![Accident Image] | Accident |
The images are categorized into "Accident" and "No Accident," which helps train the model to distinguish between accident scenes and normal traffic conditions.
Model Architecture
Our model employs a Vision Transformer (ViT) architecture, which is well-suited for image classification tasks. The key components of the model include:
- Input Layer: Accepts images resized to a specified resolution.
- Transformer Encoder Layers: Extract features through self-attention mechanisms, capturing spatial relationships.
- Feedforward Neural Networks: Process the features and classify them into accident-related categories.
- Output Layer: Provides the final classification probabilities for "Accident" and "No Accident."
Instructions for Running the Training Job
To run the training job, follow these steps:
- Clone the repository:
1git clone https://github.com/yourusername/accident-detection.git
2cd accident-detection
3
4
vit-accident-image
This model is a fine-tuned version of
google/vit-base-patch16-224-in21k on the accident classification dataset.
It achieves the following results on the evaluation set:
- Loss: 0.2027
- Accuracy: 0.93
- F1: 0.9301
Model description
label 0 : non-accident , label 1 : accident-detected
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 10
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 |
|---|
| 0.3546 | 2.0 | 100 | 0.2327 | 0.9184 | 0.9184 |
| 0.1654 | 4.0 | 200 | 0.2075 | 0.9388 | 0.9388 |
| 0.0146 | 6.0 | 300 | 0.2497 | 0.9388 | 0.9387 |
| 0.0317 | 8.0 | 400 | 0.2179 | 0.9286 | 0.9285 |
| 0.0192 | 10.0 | 500 | 0.2255 | 0.9286 | 0.9286 |
Framework versions
- Transformers 4.30.0
- Pytorch 2.2.1+cu121
- Datasets 2.19.1
- Tokenizers 0.13.3
