This document describes the code usage published with the paper ”Continuous-Time Trajectory Estimation From Noisy Camera Poses Using Cubic Bézier Curves”. The code is available at https://github.com/GSORF/VisualOdometry2Bezier.
The code was developed with the python 3.9 interpreter. Necessary additional packages to run the code are:
- matplotlib version 3.3.2 or higher
- numpy version 1.19.4 or higher
The folder \measurementData contains example measurement files, which can be used in this demonstration. To run the example execute the python file plot3D.py.
This is the main file of the code demonstration, in which a measurement file is loaded, the algorithm for optimized control point estimation called and the results are plotted, as in Fig. 1 and Fig. 2. In the file the default folder path of the measurements is defined and the desired file can be selected during the run of the code. Alternatively or if an error occurs, the file path can be setted directly e.g. measurementData/straightLine3D_3D_WNOS_P_10_NoiseS_5_001.txt. In the array plotBool at the end of the file the plot can be configured. By setting the boolean values at the specific positions to true, the following objects are contained in the plot:
- 0: camera translations
- 1: control polygon
- 2: control points
- 3: Bézier points
- 4: ground truth
- 5: camera coordinate systems
- 6: optimized control points
- 7: optimized curve
The algorithm for optimized control point estimation is called by the function estOptControlPoints. The input data is:
- non-noisy translation of the first measurement point
- non-noisy translation of the last measurement point
- all camera poses
- all absolute velocities
- all measurement time stamps
The function returns next to the optimized control points the median absolute deviation in every component of both control points. It is recommenced to use the algorithm in the debugger to follow its functionality step by step together with in line comments and the description of the paper.
The example measurement files in the folder \measurementData, like parable3D_WNOTRS_3D_P_10_NoiseT_0.1_NoiseR_5_NoiseS_1_001.txt, are named according to the following conventions: The first substring shows the trajectories shape e.g. parable3D. The second part symbolises the kind of noise on the measurement data, like WNOTRS stands for white noise on translation, rotation and velocity. In the next substring the dimension of noise is expressed. In case of 3D the translation noise is on every component of the translation and the rotation matrix is rotated around all three axis. In 2D the translation noise is only added on the x any y component and the camera coordinate frame is only rotated around the up vector. P_10 shows that in the measurement ten points are included. The numbers after the single noise strings NoiseT, NoiseR and NoiseS represent the standard deviation of noise on translation, rotation and velocity. The last part of the file name is a simple index.
The content of the measurement files is structured as follows. In the first N lines the measurement values for every point are listed. The fist number is the current time, whereby every measurement starts at 0s. The second value is the parameter u. The absolute velocity is noted next.
At last the camera pose is listed comma separated line by line. From file line N +1 until the end the true Bézier and control points are stored. With this information the ground truth can be calculated and plotted.