ModelPrésentéExterne

Accident Detection Using Dashcam

This repository contains model for predicting vehicle collisions using dashcam footage, developed for the Nexar Dashcam Collision Prediction Challenge. The model achieved a 11th place finish on the leaderboard with a...

Sourcehuggingface.co/jatinmehra/Accident-Detection-using-Dashcam/tree/a36a28966edef57ae4ce972c355c96c141288138

TypeOtherModèle de baseMCG-NJU/videomae-large-finetuned-kineticsVersionRevision a36a289BClasse

Source du modèle

Type: Other
Taille du fichier: 1,13 GB
Modèle de base: MCG-NJU/videomae-large-finetuned-kinetics
Version: vRevision a36a289
Mis à jour: 26 sept. 2026
Téléchargements: 122
Source: Hugging Face

Description de la source

🚗 VideoMAE-2 for Dashcam Collision Prediction

For Training code - GitHub

🧠 Model Overview

Architecture: VideoMAE-2 Large fine-tuned for binary classification (collision/near-miss vs. normal driving).
Feature Extraction: Utilized TimeSformer for preprocessing input frames.
Input: 16 frames per video, each resized to 224x224 pixels.
Output: Probability score indicating the likelihood of a collision or near-miss event.

📁 Dataset

The model was trained on the Nexar Collision Prediction Dataset.

750 non-collision videos
400 collision videos
350 near-miss videos arXiv

Each video is annotated with:

Event Type: Collision, near-miss, or normal driving
Event Time: Timestamp of the (near-)collision
Alert Time: Earliest time the event could be predicted.

For more details, refer to the dataset paper.

🛠️ Preprocessing Pipeline

Frame Extraction: Sampled 16 frames per video, focusing on the interval around the alert time.
Feature Extraction: Applied TimeSformer feature extractor to obtain pixel values.
Data Augmentation: Implemented transformations such as horizontal flip, rotation, color jitter, and resized cropping.
Normalization: Used ImageNet mean and standard deviation for normalization.

🏋️ Training Details

Framework: PyTorch with Hugging Face Transformers and Trainer API.
Training Configuration:
- Batch Size: 4
- Epochs: 15
- Learning Rate: 3e-5
- Weight Decay: 0.01
- Evaluation Strategy: Per epoch
- Metric for Best Model: Average Precision
Hardware: Trained on 2x NVIDIA T4 GPUs (~4.5 hours)

📊 Evaluation Metrics

The model's performance was evaluated using Mean Average Precision (mAP) across different time-to-accident intervals:

500ms
1000ms
1500ms

The final score is the mean of the Average Precision (AP) values at these intervals, emphasizing early and accurate collision predictions

📚 Citation

If you use this model or dataset, please cite:

@misc{nexar2025dashcamcollisionprediction,
  title={Nexar Dashcam Collision Prediction Dataset and Challenge},
  author={Daniel C. Moura and Shizhan Zhu and Orly Zvitia},
  year={2025},
  eprint={2503.03848},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2503.03848}
}

Sources

1 source

Hugging FacePrincipalVersion a36a28966edef57ae4ce972c355c96c141288138

huggingface.co/jatinmehra/Accident-Detection-using-Dashcam/tree/a36a28966edef57ae4ce972c355c96c141288138

Vérifié 28 août

Artefacts du modèle

1 artefact

model.safetensors

safetensors · 1,13 GB · SHA-256 bcfe7c98d033…d38f · Hugging Face

Télécharger

Extraits de sources

2 extraits

Readme28 aoûthuggingface.co/jatinmehra/Accident-Detection-using-Dashcam

--- license: mit metrics: - precision base_model: - MCG-NJU/videomae-large-finetuned-kinetics pipeline_tag: video-classification library_name: transformers new_version: jatinmehra/Accident-Detection-using-Dashcam datasets: - nexar-ai/nexar_collision_prediction tags: - vision --- # 🚗 VideoMAE-2 for Dashcam Collision Prediction This repository contains model for predicting vehicle collisions using dashcam footage, developed for the [Nexar Dashcam Collision Prediction Challenge](https://www.kaggle.com/competitions/nexar-collision-prediction). The model achieved a **11th place** finish on the leaderboard with a score of **0.80** mean Average Precision (mAP) For Training code - [GitHub](https://github.com/Jatin-Mehra119/-Nexar-Dashcam-Crash-Prediction-Challenge) ---------- ## 🧠 Model Overview - **Architecture**: [VideoMAE-2 Large](https://huggingface.co/MCG-NJU/videomae-large-finetuned-kinetics) fine-tuned for binary classification (collision/near-miss vs. normal driving). - **Feature Extraction**: Utilized [TimeSformer](https://huggingface.co/facebook/timesformer-base-finetuned-k400) for preprocessing input frames. - **Input**: 16 frames per video, each resized to 224x224 pixels. - **Output**: Probability score indicating the likelihood of a collision or near-miss event. ---------- ## 📁 Dataset The model was trained on the [Nexar Collision Prediction Dataset](https://huggingface.co/datasets/nexar-ai/nexar_collision_prediction). - 750 non-collision videos - 400 collision videos - 350 near-miss videos [arXiv](https://arxiv.org/html/2503.03848v1?utm_source=chatgpt.com) Each video is annotated with: - **Event Type**: Collision, near-miss, or normal driving - **Event Time**: Timestamp of the (near-)collision - **Alert Time**: Earliest time the event could be predicted. For more details, refer to the [dataset paper](https://arxiv.org/abs/2503.03848). ---------- ## 🛠️ Preprocessing Pipeline 1. **Frame Extraction**: Sampled 16 frames per video, focusing on the interval around the alert time. 2. **Feature Extraction**: Applied TimeSformer feature extractor to obtain pixel values. 3. **Data Augmentation**: Implemented transformations such as horizontal flip, rotation, color jitter, and resized cropping. 4. **Normalization**: Used ImageNet mean and standard deviation for normalization. ---------- ## 🏋️ Training Details - **Framework**:...

Note28 aoûthuggingface.co/jatinmehra/Accident-Detection-using-Dashcam

Source context: 122 downloads · 1 likes · Pipeline video-classification · Library transformers · Repo jatinmehra/Accident-Detection-using-Dashcam