Graphs

[shadow1runner]

• Difference between Kalman-filtered FOE, raw FOE, and Kalman-filtered FOE with descend detection.
  • compare to Kalman with Sebastian (which used a fitted polynomial, even though the code didn't show any of these signs)

# OBSOLETE -> see below

[shadow1runner]

• Table: System performance

as well as graph analogous to Fig. 9 in

@inproceedings{al2016monocular, title={Monocular vision-based obstacle detection/avoidance for unmanned aerial vehicles}, author={Al-Kaff, Abdulla and Meng, Qinggang and Mart{\'\i}n, David and de la Escalera, Arturo and Armingol, Jos{\'e} Mar{\'\i}a}, booktitle={2016 IEEE Intelligent Vehicles Symposium (IV)}, pages={92--97}, year={2016}, organization={IEEE} }

# DONE (graph only)

[shadow1runner]

• ground obstacle detection
  • descend detection?

# DONE

[shadow1runner]

• difference between distortion enabled/disabled, a good example is the balconyCrash (/home/helli/d/m/qgroundcontrol/src/CollisionAvoidance/opticalflow/res/boscam/fisheye/longFisheye/90degDown/balconyCrash.avi) -> also include the optical flow field generated which shows the radial nature of the non-undistorted frame

# DONE

[shadow1runner]

• Impact of CollisionLevel degradation over time

# DONE

[shadow1runner]

• Comparison to perspective camera in sythetic video sequences using the omnidirectional camera renderer blender plugin (here and here)

[shadow1runner]

• Comparison of KF (cf. #55) to Sebastian's settings, compare FOE movement and discuss Proposal: implement two KF instances and let both do their magic - compare this movement, easier than merging Scala and C++

Branch graph/kf and QGroundControl_kalman.ini shows results have been drawn, it would be best to include 1e5/kalman_balconyCrash.pdf shows ok-ish values which can be used.

# DONE

[shadow1runner]

• Difference between divergence obtained by
  • an affine model
  • the 'differential' method in both time and value, cf. Figure 8.8 & 8.9 in Sebastian's thesis

# PROGRESS: There's hardly any correlation between those two methods (0.33 in figures), the affine model showed better results wrt. almost all unit tests - even though the divergence thresholds are fundamentally different (as they don't correlate to the differential method as mentioned before)

#TODO: Redo performance measurements with new affine model

[shadow1runner]

• Influence of rotation in the evaluation (gimbal etc.)
  • bottom-mounted static mounted camera and comparisons with this footage?

-> think about a meaningful possibility to compare these two methods, maybe also compare it to Stabiner, 2014

[shadow1runner]

• lateral collision detection

[shadow1runner]

• determine optimal number of vectors in affine differgence wrt.
  • accuracy (in comparison to use all vectors in a patch)
  • performance (the less the better, I assume?)