Anticipating Accidents in Dashcam Videos (ACCV 2016)
Switch branches/tags
Nothing to show
Clone or download
Fu-Hsiang Chan
Fu-Hsiang Chan support tensorflow=1.6
Latest commit ef3b65e Aug 24, 2018
Permalink
Failed to load latest commit information.
dataset first commit Aug 2, 2017
demo_model first commit Aug 2, 2017
img first commit Jul 30, 2017
README.md support tensorflow=1.6 Aug 24, 2018
accident.py support tensorflow=1.6 Aug 24, 2018

README.md

Anticipating Accidents in Dashcam Videos

By Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, Min Sun.

Introduction

Anticipating Accidents in Dashcam Videos is initially described in a ACCV 2016 paper. We propose a Dynamic-Spatial-Attention (DSA) Recurrent Neural Network (RNN) for anticipating accidents in dashcam videos.

Requirements

Tensoflow 1.x
Opencv 2.4.9
Matplotlib
Numpy

Model Flowchart

Alt text

Dataset & Features

  • Dataset : link (Download the file and put it in "datatset/videos" folder.)

  • CNN features : link (Download the file and put it in "dataset/features" folder.)

  • Annotation : link

If you need the ground truth of object bounding box and accident location, you can download it.

The format of annotation:

<image name, track_ID, class , x1, y1, x2, y2, 0/1 (no accident/ has accident)>

Usage

Run Demo

python accident.py --model ./demo_model/demo_model

Training

python accident.py --mode train --gpu gpu_id

Testing

python accident.py --mode test --model model_path --gpu gpu_id

Citing

Please cite this paper in your publications if you use this code for your research:

@inproceedings{chan2016anticipating,
    title={Anticipating accidents in dashcam videos},
    author={Chan, Fu-Hsiang and Chen, Yu-Ting and Xiang, Yu and Sun, Min},
    booktitle={Asian Conference on Computer Vision},
    pages={136--153},
    year={2016},
    organization={Springer}
}