This code/implementation is available for research purposes. If you are using this code/data for your work, please cite the following paper:

Huynh, Manh, and Gita Alaghband. "GPRAR: graph convolutional network based pose reconstruction and action recognition for human trajectory prediction." arXiv preprint arXiv:2103.14113 (2021). In Proceeding of BMVC 2021. https://arxiv.org/pdf/2103.14113.pdf

Key ideas:

• GPRAR is a graph convolutional network based pose reconstruction and action recognition for human trajectory prediction.
• GPRAR learns robust features: human poses and actions, under noisy scenarios.
• By design, GPRAR consists of two modules: PRAR (Pose Reconstruction and Action Recognition) and FA (Feature Aggregator).

Experiments on JAAD dataset

1. Generate train/val data
a. Read PRE_PROCESS.md for instructions extracting features.
b. Generate/val data of JAAD and Kinetics for reconstruction task:

$ python data_procesing/reconstruction/generate_data_jaad.py 
$ python data_procesing/reconstruction/generate_kinetics_jaad.py 

c. Generate train/val data for prediction task with different observation types obs_type: noisy, impute, gt :

$  python data_processing/prediction/generate_data_jaad.py --obs_type noisy
$  python data_processing/prediction/generate_data_jaad.py --obs_type impute
$  python data_processing/prediction/generate_data_jaad.py --obs_type gt