About this Project

This structure from motion simulator is created as a workspace to test classification and weighting methods on SLAM systems in dynamic environment. It uses an internal (non-visual) representation of dynamic and static features which are assumed to be tracked without any mismatch but are affected by noise on the image detection. Unlike common bundle adjustment, this simulator uses a modified costfunction inside ceres to classifiy and weight features based on their reprojection error. This simulator was used in a first stage in my master thesis about "Visual SLAM for Dynamic Environemnts" at the V4RL at ETH Zurich.

Building

This project is built and tested with g++ (Debian 7.3.0-16) 7.3.0 (should work on other unix systems aswell).

Dependencies Simulator

• c++17
• Ceres
• OpenCV
• OpenCV contrib modules (OpenCV_SFM)
• Eigen
• Boost (program_options)

Dependencies Plotting

• pandas
• Numpy
• scipy
• matplotlib
• seaborn

Build with CMake

mkdir build && cd build
cmake ..
make -j4

Run the Simulation

For a quick test-run follow Summary.

Generate Dynamic/Static Features and Cameraposes

Run the genpoints.py in the scripts folder to generate feature points with the following arguments:

python genpoints.py <outputfolder> <numstaticfeatures> <numdynamicfeatures> <numframes> <movement> <startstatic> <stopstatic>

For example to generate a dataset of 30 static points and 20 dynamic (which move in random directions) for 120 frames in the dataset D0 (folder has to exist):

python genpoints.py ../data/D0 30 20 120 false 1 1

Run sfmsimulator

Just call the executable from the build folder (important!) with the following arguments.

cd build
./bin/sfmsimulation -o <outputfolder> -c <classifier1> -c <classifier2> -t <dataset1> -t <dataset2> ...

• -c 0 (= no classifier)
• -c 1 (= no dependency classifier)
• -c 2 (= dependecy "dep3" classifier)

Plot data

These python scripts use pandas, seaborn, scipy, numpy and matplotlib to plot different data to analyse and evaluate the method mentioned above.

First, one has to change the filepaths to the result folder. In this result folder the output of the simulator should lie in a seperate folder for each run. Eg: <path>/result/run0.

Look around in the scripts folder to plot frames of the trajectory or other data for the evaluation.

Plot Average Weight of Features

python plot_average.py <run> <dataset> <numberofdynamicpoints>

eg:

python plot_average.py run0 dataset0 20

Plot Camera Trajectory Accuracy

python plot_camera_accuracy.py <dataset>

Summary

This example test-run for all classifiers prints the average weights for each classifier for each frame:

mkdir build && cd build
cmake ..
make -j4
mkdir -p ../data/D0
mkdir -p ../results/R0
python ../scripts/genpoints.py ../data/D0 50 20 60 false 1 1
./bin/sfmsimulation -o ../results/R0/ -c 0 -c 1 -c 2 -t D0
python ../scripts/plot_average.py R0 D0 20