This project proposes a locaNet for multi-robot localization by predicting the pixel position of the robot center and its distance from the camera, which can be transformed to inertial 3D relative positions using the intrinsic parameter. The implementation contains several code projects: 1) Blender code for synthetic dataset generation; 2) TensorFlow-based network of the locaNet and training; 3) Onboard deep network code on the AIdeck; 4) Quadrotor control code based on Crazyflie firmware.
.
├── aideck-locaNet # deep network on the AI edge chip - GAP8
├── dataset #
├── AIdeck-dataset # flight images and relative positions
├── Other files # synthetic images rendered by Blender
├── Otherfiles # neural network, training, and testing
└── README.md
Paper: PDF on arXiv.
Video: Real-world flight on Youtube.
- locaNet & locaAIdeck (Python 3.6, pip=20.1.1, tensorflow==2.1.2, h5py==2.10.0)
- dataset (Blender, numpy, csv, statistics)
$ git clone https://github.com/shushuai3/deepMulti-robot.git