Human Detection with Ros

• This repo illustrates how to use Human Detection modul with usb_cam by Ros
• We need Ros, pytorch, Cvbridge, etc..
• requirments must be installed (see Reference).
• usb_cam launch file is in face-recognition-ros repo see that.

Ros

• We will skip step of basic Ros settigs.
• Architecture
  • Publish usb_cam images with Ros

    roslaunch usb_cam usb_cam-test.launch
    

  • Subscribe image topics in code, and convert it to Opencv_images with CVbridge
  • with Opencv images do Human Detection
  • Publish result bbox, score, num of human.

Object Detection

• Extract bbox(bounding box) and class
• input : image data
• output : [xmin, ymin, xmax, ymax, box confidence, class, class confidence(score)]

Human Detection

• In Object Detection only detect human.
• Using transfer learning
• Trained by AVA dataset

Dataset

AVA dataset

• In Youtube video, doing bbox annotation
• you can download dataset here [https://github.com/DoranLyong/AVA-Dataset-Processing-for-Person-Detection]

  • How to download

    • clone github repository

    git clone https://github.com/DoranLyong/AVA-Dataset-Processing-for-Person-Detection.git
    cd AVA_Dataset_Processing-for-Person_Detection.git
    mkdir dataset

    • Get in[https://research.google.com/ava/download.html#ava_kinetics_download]download ava_v2.2.zip and Unzip the file into dataset
    • setup.sh

    bash setup.sh

    • Requiremnet

    pip install -r requiremnets.txt

    • Download YouTube Video and Image Frames

    python ava_youtube_download.py
    python cut_frames_from_video.py

    • Detection labels
      • AVA dataset has 430 video files(235 : training / 64 : validation / 131 : test)
      • Label have data about human localization & action recognition
      • Get YOLO format

      python cvt_annotation_format_csv_to_txt.py
      python label_test.py

COCO dataset

• You can download in here https://cocodataset.org/#home
• 80 object classes but we use only human class.
• coco data format

YOLO -> COCO format

• To train YOLOX, we need to convert dataset YOLO format to COCO format.
• https://github.com/RapidAI/YOLO2COCO

  YOLOV5
  ├── classes.txt
  ├── xxxx
  │   ├── images
  │   └── labels
  ├── train.txt
  └── val.txt
  

• In dataset make dir like above.
• classes.txt file has only human class
• train.txt and val.txt have to made by self.
• Run

  python yolov5_2_coco.py --dir_path dataset/YOLOV5

YOLOX-nano for human detection

1. Clone YOLOX repository

git clone https://github.com/Megvii-BaseDetection/YOLOX.git

2. Prepare Dataset

AVA dataset

• move yolo2coco dataset to YOLOX/dataset

3. Training YOLOX nano

• In YoloX github download weight of YOLOX-nano
• Set weight file to YOLOX
• Traning

cd YOLOX
python tools/train.py -f exp/default/nano.py -d 1 -b 8 --fp16 -c yolox_nano.pth

• -d : num of devices (# of gpus) -b : num of batch (recommended : 8 times of device)
• -c : pth file for transfer learning

How to use (you must launch 'usb_cam')

- Run model

rosrun yolox inference.py

- Publish msg to activate

rostopic pub --once /human_detection_msg std_msgs/String "data: 'On'"

Reference

• YOLOX-nano link: https://github.com/Megvii-BaseDetection/YOLOX.git

  git clone https://github.com/Megvii-BaseDetection/YOLOX.git
  

• Usb_cam link: https://github.com/ros-drivers/usb_cam.git

  git clone https://github.com/ros-drivers/usb_cam.git
  

• yolo2coco link: https://github.com/RapidAI/YOLO2COCO

  git clone https://github.com/RapidAI/YOLO2COCO