PyDNet based Collision Avoidance

This repository contains the source code of pydnet-based collision avoidance implemented on a DJI-Tello drone, as proposed in the paper "Towards real-time unsupervised monocular depth estimation on CPU", IROS 2018.

For more details on the research: arXiv

Requirements

• Tensorflow 1.8 (recomended)
• python packages such as opencv, matplotlib
• PyDNet Framework
• Monodepth Framework

Run PyDNet on Tello feed

To run pydnet, just launch

python3 tello_pydnet_interface.py --checkpoint_dir /checkpoint/IROS18/pydnet --resolution [1,2,3]

Please note that the velocity commands have been commented out. You could either uncomment them or create your own navigation algorithm.

Navigation Algorithm

Navigation towards the region of most depth. Yawing action performed till the maximum depth region is in and around 20% from the frame center.

Train PyDNet from Scratch

Requirements

• monodepth (https://github.com/mrharicot/monodepth) framework by Clément Godard

After you have cloned the monodepth repository, add to it the scripts contained in training_code folder from this repository (you have to replace the original monodepth_model.py script). Then you can train pydnet inside monodepth framework.

Evaluate PyDNet on Eigen split

To get results on the Eigen split, just run

python3 experiments.py --datapath PATH_TO_KITTI --filenames PATH_TO_FILELIST --checkpoint_dir /checkpoint/IROS18/pydnet --resolution [1,2,3]

This script generates disparity.npy, that can be evaluated using the evaluation tools by Clément Godard