Skip to content

Latest commit

 

History

History
129 lines (100 loc) · 5.66 KB

README.md

File metadata and controls

129 lines (100 loc) · 5.66 KB
Raw

Leveraging-Trajectory-Prediction-for-Pedestrian-Video-Anomaly-Detection

Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du

Clone Repo

git clone --recurse-submodules https://github.com/akanuasiegbu/Leveraging-Trajectory-Prediction-for-Pedestrian-Video-Anomaly-Detection.git

Installation

  • scipy==1.4.1
  • matplotlib==3.3.1
  • Pillow==7.2.0
  • scikit_learn==0.23.2
  • opencv-python==4.4.0.42
  • jupyter
  • jupyterthemes==0.20.0
  • hyperas==0.4.1
  • pandas==1.1.2
  • seaborn==0.11.0
  • tensorflow_addons==0.11.2
  • tensorflow_datasets
  • wandb==0.10.12
  • more_itertools==8.8.0

You can also use docker with 'docker/Dockerfile'. Note that I set the PYTHONPATH inside docker file would need to adjust that path "ENV PYTHONPATH "/mnt/roahm/users/akanu/projects/anomalous_pred/custom_functions:/home/akanu".

Step 1: Download Dataset

  • The extracted bounding box trajectories for Avenue and ShanghaiTech with the anomaly labels appended can be found here .
  • To want to recreate the input bounding box trajectory

Step 2: Training

We used two two models for our experiments Long Short Term Memory (LSTM) Model and BiTrap model.

Training LSTM Models

  • Users can train their LSTM models on Avenue and ShanghaiTech
    • Training Avenue: python models.py
      • In config.py change ```hyparams['input_seq'] and hyparams['pred_seq'] to match input/output trajectory length
    • Training ShanghaiTech: python models.py
      • In config.py change ```hyparams['input_seq'] and hyparams['pred_seq'] to match input/output trajectory length

Training BiTrap Model

  • For training BiTrap models refer forked repo here.

Train on Avenue Dataset

cd bitrap_leveraging
python tools/train.py --config_file configs/avenue.yml

Train on ShanghaiTech Dataset

cd bitrap_leveraging
python  tools/train.py --config_file configs/st.yml

To train/inferece on CPU or GPU, simply add DEVICE='cpu' or DEVICE='cuda'. By default we use GPU for both training and inferencing.

Note that you must set the input and output lengths to be the same in YML file used (INPUT_LEN and PRED_LEN) and bitrap_leveraging/datasets/config_for_my_data.py (input_seq and pred_seq)

Step 3: Inference

Pretrained BiTrap Model:

Trained BiTrap models for Avenue and ShanghiTech can be found here

Pretrained LSTM Models:

Trained LSTM models for Avenue and ShanghiTech can be found here

LSTM Inference

We do not explictly save the LSTM trajectory outputs into a file (such as pkl). Therefore the inference and the AUC calcution step for the LSTM model are performed simultaneously. Please refer to LSTM AUC Calcuation section shown below.

BiTrap Inference

To obtain BiTrap PKL files containing the pedestrain trajectory use commands below. Test on Avenue dataset:

cd bitrap_leveraging
python tools/test.py --config_file configs/avenue.yml CKPT_DIR **DIR_TO_CKPT**

Test on ShanghaiTech dataset:

cd bitrap_leveraging
python tools/test.py --config_file configs/st.yml CKPT_DIR **DIR_TO_CKPT**
PKL Files

BiTrap pkl files can be found here.

  • Download the output_bitrap folder which contains the pkl file folders for Avenue and ShanghiTech dataset.
  • Naming convention: in_3_out_3_K_1 means input trajectory and output trajectory is set to 3. And K=1 means using Bitrap as unimodal.

Step 4: AUC Caluation

BiTrap AUC Calcuation

  • In experiments_code/run_bitrap_auc.py make sure exp['model_name']='bitrap'
  • Then set hyparams['input_seq'] and hyparams['pred_seq'] to desired length
  • Set hyparams['metric'] to either 'giou', 'l2' ,or 'iou'
  • Set hyparams['errortype'] to either 'error_summed' or 'error_flattened'
  • To run change load_pkl_file varible located in run_auc_bitrap.py to desired location
    • Then use python run_bitrap_auc.py

LSTM AUC Calcuation

  • In experiments_code/run_lstm_auc.py make sure exp['model_name']='lstm_network'
  • Then set hyparams['input_seq'] and hyparams['pred_seq'] to desired length
  • Set hyparams['metric'] to either 'giou', 'l2' ,or 'iou'
  • Set hyparams['errortype'] to either 'error_summed' or 'error_flattened'
  • To run change pretrained_model_loc varible located in run_lstm_auc.py to desired location of pretrained lstm model
    • Then use python run_lstm_auc.py

If you want to run multiple LSTM/AUC refer to run_quick.py

Citation

If you found repo useful, feel free to cite.

@INPROCEEDINGS{9660004,
  author={Kanu-Asiegbu, Asiegbu Miracle and Vasudevan, Ram and Du, Xiaoxiao},
  booktitle={2021 IEEE Symposium Series on Computational Intelligence (SSCI)}, 
  title={Leveraging Trajectory Prediction for Pedestrian Video Anomaly Detection}, 
  year={2021},
  volume={},
  number={},
  pages={01-08},
  doi={10.1109/SSCI50451.2021.9660004}}