Skip to content
master
Switch branches/tags
Go to file
Code

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 

mht-paf

Video results of all sequences and trained models can be downloaded in Google drive.

Multiple Human Tracking using Multi-Cues including Primitive Action Features
Hitoshi Nishimura, Kazuyuki Tasaka, Yasutomo Kawanishi, Hiroshi Murase
https://arxiv.org/abs/1909.08171

MCF MCF MHT-PAF MHT-PAF

Requirements

  • NVIDIA driver (>= 410.48)
  • Docker (>= 17.12.0)
  • nvidia-docker2
  • docker-compose (>= 1.21.0)

Setup

Clone this repository:

git clone --recursive https://github.com/hitottiez/mht-paf.git

Build a docker image using docker-compose:

cd mht-paf
cp env.default .env
docker-compose build

Start a docker container:

docker-compose up -d
docker-compose ps
        Name             Command    State          Ports        
----------------------------------------------------------------
mht-paf_deepsort_1      /bin/bash   Up                          
mht-paf_mcf-tracker_1   /bin/bash   Up                          
mht-paf_tsn_1           /bin/bash   Up      0.0.0.0:8888->80/tcp

Login/Logout docker container

Login each container:

docker-compose exec <deepsort or mcf-tracker or tsn> bash

or, follow the README.md in each repository.

Logout each container:

exit

Preparing dataset

Note: Dataset is assumed to be in the dataset directory. If you change the dataset directory, change DATASET_DIR in the .env file.

Download all feature files from Google drive.

Directory structure:

dataset
└── images
     ├── 1.1.1
     │   └── feature_results
     │       ├── cnn.txt
     │       ├── det.txt
     │       └── fusion_tsn.txt
     ├── 1.1.2
     ├── 1.1.3
     ...
     ├── 2.2.10
     └── 2.2.11

Download all videos and labels (SingleActionTrackingLabels, MultiActionLabels) from Okutama-Action dataset.

The labels are set in the images directory:

dataset
├── images
│
├── labels # SingleActionTrackingLabels
│   ├── test
│   │   ├── 1.1.8.txt
│   │   ├── 1.1.9.txt
│   │   ...
│   │   └── 2.2.10.txt
│   └── train
│        ├── 1.1.1.txt
│        ├── 1.1.2.txt
│        ...
│        └── 2.2.11.txt
│
└── multi_labels # as same as MultiActionLabels labels

Convert videos to JPG images using ffmpeg:

docker-compose exec deepsort bash

# in deepsort/mcf-tracker container
ffmpeg -i <path/to/download>/Train-Set/Drone1/Morning/1.1.1.mov  -vcodec mjpeg -start_number 0 <path/to/dataset>/images/1.1.1/%d.jpg
ffmpeg -i <path/to/download>/Train-Set/Drone1/Morning/1.1.10.mp4  -vcodec mjpeg -start_number 0 <path/to/dataset>/images/1.1.10/%d.jpg
...
ffmpeg -i <path/to/download>/Train-Set/Drone2/Noon/2.2.11.mp4  -vcodec mjpeg -start_number 0 <path/to/dataset>/images/2.2.11/%d.jpg &

Running human tracking and action recognition

We provide two choices, deepsort and mcf-tracker.

Evaluation

Refer deepsort.

About

Multiple Human Tracking using Multi-Cues including Primitive Action Features (MHT-PAF)

Resources

License

Releases

No releases published

Packages

No packages published