TTCDist: Fast Distance Estimation From an Active Monocular Camera Using Time-to-Contact

This repository contains code for TTCDist: Fast Distance Estimation From an Active Monocular Camera Using Time-to-Contact by Levi Burner, Nitin J. Sanket, Cornelia Fermuller, and Yiannis Aloimonos.

You can find a pdf of the paper here.

Setup

We use standard scientific Python (>=3.8) libraries like NumPy, SciPy, Matplotlib, and Numba. Make sure these are installed.

The python bindings of the official apriltag library are needed to use the --april flag in any of the scripts below.

The Python package pyrealsense2 is needed to run the algorithms in realtime with a RealSense D435i camera. The following links may be helpful when tring to install pyrealsense2 on Ubuntu 22.04 and newer. Older versions of Ubuntu should use the standard instructions.

• https://github.com/IntelRealSense/librealsense/blob/master/doc/installation.md
• https://github.com/IntelRealSense/librealsense/tree/master/wrappers/python

Running from a recording in a folder

Data

We provide the data for the 10 sequences considered in TTCDist here.

Download and extract to TTCDist/recordings_21_11_12

Running on one sequence

python3 ttc_depth_from_folder.py --dir recordings_21_11_12/record_000000 --visualize

Benchmark on a single sequence

Note: On the first run, the Numba functions will be compiled which will significantly affect the runtime. However, when run a second time, a compilation cache will be used instead.

python3 ttc_depth_from_folder.py --dir recordings_21_11_12/record_000000 --bench

Benchmark on all sequences

Note: On the first run, the Numba functions will be compiled which will significantly affect the runtime. However, when run a second time, a compilation cache will be used instead.

python3 ttc_depth_from_folder.py --dir recordings_21_11_12 --all --bench

Calculate Average Trajectory Error

We provide the data produced by all methods on the sequences in TTCDist here.

Download and extract to TTCDist/results

Then run the following to regenerate the results for the Tau, Phi, and AprilTag methods and calculate error.

python3 ttc_depth_from_folder.py --dir recordings_21_11_12 --all --save --april
python3 ttc_depth_calc_error.py --dir results --calc --latex

To use the --april flag the python bindings for the official apriltag package must be installed from here.

If you wish to reevaluate ROVIO and VINS-Mono on the recorded sequences the scripts in the ros/ folder may be

Running in realtime with an Intel Realsense D435i

Without apriltag package

Assuming the pyrealsense2 package is installed just run the following

python3 ttc_depth_realsense.py --visualize

A live output will run and the Z distance between the camera and the object will be shown. The top number is from the Tau constraint and the bottom is from the Phi constraint.

The r key can be used to reset the patch being tracked. The q key will close the application.

With apriltag ground truth

If the python bindings for the official apriltag package are installed from here then the following will work

python3 ttc_depth_realsense.py --visualize --april

in another terminal run

python3 ttc_depth_plot_live.py

A live plot will show. The top row shows the X, Y, and Z positions of the tracked object. x_hat is from the Tau Constraint. x_gt is from the AprilTag, and phi_x_hat is from the Phi Constraint.

Running the RoboMaster experiments

Install the RoboMaster-SDK Python package. Make sure your RoboMaster robot works with some of the examples included in the SDK. Then to run the experiments from TTCDist use the command:

python3 ttc_depth_robomaster.py --visualize --b 2.0 --efference

Remove the --efference flag to use measured acceleration as described in the paper.

Acknowledgements

Part of the rpg_trajectory_evaluation toolbox is reproduced in rpg_align_trajectory/ and is available under its respective license.