PoseRisk: Pose-based Risk Scoring

Introduction

This repository is the implementation of an automatic human hazard detection model based on a human posture. This project has collaborated with SNU HIS LAB.

Install guidelines

We recommend you to use an Anaconda virtual environment. Install PyTorch >= 1.2 according to your GPU driver and Python >= 3.7.2.
Install the requirements using conda:

sh script/install_conda.sh

Preparations

We used Simple Multi Person Tracker for video human detection, and SPIN for human pose estimation. If these are not in lib directory after installation, put these two repository in lib directory.

${ROOT}  
|-- lib  
|   |-- [multi-person-traker](https://github.com/mkocabas/multi-person-tracker)
|   |-- [SPIN](https://github.com/nkolot/SPIN)

The data directory structure should follow the below hierarchy.

${ROOT}  
|-- data  
|   |-- base_data  
|   |   |-- human_models  
|   |   |-- spin_data

• Download basicModel_f_lbs_10_207_0_v1.0.0.pkl, basicModel_m_lbs_10_207_0_v1.0.0.pkl, and basicModel_neutral_lbs_10_207_0_v1.0.0.pkl from here (female & male) and here (neutral) to ${ROOT}/data/base_data/human_models.
• Download SPIN data from here. Please reference Fetch data section in the repository.

Run code

We provide REBA and RULA score method. You should put some additional information for score estimation. Please refer example/additional_information.json.

Most of the evaluation criteria referred from: https://kosha.or.kr/kosha/business/musculoskeletal_c_d.do
for REBA,

• Legs_bilateral_weight_bearing/walking: (1-2) If balanced of two legs 1, if not 2
• Sitting: (0-1) If seated 1, if not 0
• Load/Force Score_L/R): (0-3) load amount, rapid build up of force, etc.
• Arm_supported_leaning_L/R: (0-1) If arm is supported or person is learning 1, if not 0 (IMPORTANT)
• Coupling_L/R: (0-3) well fitting handle, acceptable body part, etc.
• Activity_Score_L/R: (0-3) repeated small range actions large acitivity, etc.

for RULA,

• Arm_supported_leaning_L/R: (0-1) If arm is supported or person is learning 1, if not 0 (IMPORTANT)
• A_Muscle_use_L/R, A_Load/Force_L/R: (0-1, 0-3) load amount, rapid build up of force, etc.
• Legs_bilateral_weight_bearing: (1-2) If balanced of two legs 1, if not 2
• B_Muscle_use, B_Load/Force : (0-1, 0-3) well fitting handle, acceptable body part, etc.

Below is the running example.

Example:

cd ${ROOT}
conda activate risk
python main/run.py --type REBA,RULA --input {input video path} --info {additional information path} --output {output directory} 

Debug mode

If want to get score sequences, you should use debug option.

python main/run.py --type REBA,RULA --input {input video path} --info {additional information path} --output {output directory} --debug

If want to get pose sequences for several joints, you should add debug_joints option.

python main/run.py --type REBA,RULA --input {input video path} --info {additional information path} --output {output directory} --debug --debug_joints {joint names for debugging}

If want to get estimation result only for one frame, you should add debug_frame option.

python main/run.py --type REBA,RULA --input {input video path} --info {additional information path} --output {output directory} --debug --debug_frame {id of frame}

Reference

TBD