R-TOD: Real-Time Object Detector with Minimized End-to-End Delay for Autonomous Driving

Hardware

• Nvidia Jetson AGX Xavier
• Logitech C930e USB camera

Software

• Ubuntu 18.04 with JetPack-4.2.2
• OpenCV-3.3.1
• CUDA 10.0

More details

• http://pjreddie.com/darknet/yolo/
• https://github.com/AlexeyAB/darknet

Installation

• Clone R-TOD (Submodule: https://github.com/AveesLab/OpenCV-3.3.1)

$ git clone --recursive https://github.com/AveesLab/R-TOD

• To use On-demand Capture with OpenCV, you don't need any modification. Just build it. See OpenCV rebuild in https://github.com/AveesLab/OpenCV-3.3.1.

Compile using 'Make'

• V4L2=1: Fetch image with On-demand capture method using V4L2 ioctl without OpenCV library (0: Fetch image using OpenCV).
• ZERO_SLACK=1: Use Zero-Slack Pipeline method
• CONTENTION_FREE=1: Use Contention-Free Pipeline method
• MEASUREMENT=1: Measure delay (capture ~ display) and log to csv file (See Measurement setup).

Measurement setup

• If you build with MEASUREMENT=0, application will not stop until terminated by user.
• In src/rtod.h, you can modify measurement setup.

/* Measurement */
#define MEASUREMENT_PATH      // Directory of measurement file
#define MEASUREMENT_FILE      // Measurement file name
#define OBJ_DET_CYCLE_IDX     // Count of measurement

Usage

Original Darknet

$ ./darknet detector demo cfg/coco.data cfg weights 
      cfg: YOLO network configure file
  weights: weights file

+On-demand Capture

• If you build with V4L2=0

$ export OPENCV_QLEN=0

• If you build with V4L2=1: No setup required.

$ ./darknet detector demo cfg/coco.data cfg weights 
      cfg: YOLO network configure file
  weights: weights file

Zero-Slack Pipeline

• Zero-Slack Pipeline needs On-demand Capture.
• Build with ZERO_SLACK=1.

$ export OPENCV_QLEN=0
$ ./darknet detector rtod cfg/coco.data cfg weights
       cfg: YOLO network configure file
   weights: weights file

Contention-Free Pipeline

• Contention-Free Pipeline needs On-demand Capture.
• Build with CONTENTION_FREE=1.

$ export OPENCV_QLEN=0
$ ./darknet detector rtod cfg/coco.data cfg weights
       cfg: YOLO network configure file
   weights: weights file

Citation

The paper can be found here. For citation, please use the following Bibtex.

@misc{jang2020rtod,
      title={R-TOD: Real-Time Object Detector with Minimized End-to-End Delay for Autonomous Driving}, 
      author={Wonseok Jang and Hansaem Jeong and Kyungtae Kang and Nikil Dutt and Jong-Chan Kim},
      year={2020},
      eprint={2011.06372},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}