/CarCrash_forecasting_and_detection

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Accident Forecasting in Traffic Camera CCTV Videos

Requirements

Please use Python 2.7 and install required packages by

$ pip install -r requirements.txt

Please edit Keras configuration to load Tensorflow backend (we don't use Theano backend). A GPU is recommended.

Train/Test with original Faster R-CNN

Data preparation

Assume that you have CADP dataset at CADP_IMAGE_HOME. For example, for video with ID 0, the frames should be in CADP_IMAGE_HOME/000000/.

Generating images

There are unlabeled regions in each images, therefore for task like object detections or tracking, anomalous detection, we should generate images with those regions are masked out. To generate mask images, please modify the output path CADP_MASK_HOME in cover_crowd_far.py, and then please run:

$ python analysis/cover_crowd_far.py --anno_dir=./data/annotations

where anno_dir is the directory containing VATIC format annotations (each file contains annotations for one video).

Generating CSV annotations

Please run the following command to output csv data for training/testing.

$ python analysis/generate_csv.py --anno_dir=./data/annotations/trainval/ --csv_output=./cadp_train.csv --use_mask=True

where csv_output is the output csv files and use_mask is the flag to specify whether to use masked images.

Download pretrained models

Please download pretrained Resnet-50 model from https://github.com/fchollet/keras/tree/master/keras/applications. Assuming that you have put the pretrained Resnet model at MODEL_DIR/resnet50_weights_tf_dim_ordering_tf_kernels.h5. Please set that path as the value of cfg.base_net_weights at L25 of train_cadp_frcnn.py.

Training

After all above steps are done, please train Faster R-CNN with

$ python analysis/train_cadp_frcnn.py

Testing

After training, to measure the mAP@IoU=0.5, please run

$ python analysis/measure_map.py --path=./cadp_test.csv --parser=simple

Where path is the CSV files containing the annotations of test set and parser must be simple to parse the annotations.

Website -

Accident Forecasting Traffic Camera

Citation -

@INPROCEEDINGS{8639160,
  author={A. P. {Shah} and J. {Lamare} and T. {Nguyen-Anh} and A. {Hauptmann}},
  booktitle={2018 15th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS)}, 
  title={CADP: A Novel Dataset for CCTV Traffic Camera based Accident Analysis}, 
  year={2018},
  volume={},
  number={},
  pages={1-9},}