1. www.cranfield.ac.uk
• Reduction of ground plane error by a factor of two
compared to original algorithm.
• Reduction of mistakenly detected obstacles.
• Improved height estimation with an innovative
joined membership based on colour and disparity.
• Scene understanding approach with a good
overall accuracy and high classes recall provides
a decent scene segmentation for high-level tasks.
• 3D visualization showing stixels in context of the
laser measurements collected with the Velodyne.
Zygfryd Wieszok
Supervisor: Dr Nabil Aouf
zygfryd.wieszok@cranfield.ac.uk
Stixel based scene
understanding for
autonomous vehicles
Introduction
Autonomous vehicles are future mean of transport.
The important task in autonomous driving is scene
understanding and obstacle detection. This study
provides a scene understanding method based on
the Stixel World estimation. The method
determinates the free-space ahead of the car and
detects the obstacles. Next, the obstacles are
classified into two meaningful classes.
1. Ground plane estimation
1.2 Improved ground line estimation
1.1 Online learned colour road model
2. Improved stixel distance
Error Error
Original paper 3.847 5.708
This paper 1.770 2.600
Error Error
LS approximation 5.482 8.228
IRLS approximation 3.847 5.708
3. Improved stixel height estimation
4. Scene understanding
5. 3D visualisation
Conclusions
Vegetation and
infrastructure
Car and pedestrian
Recall Precision Recall Precision
Pixel level 81.50% 89.57% 67.68% 51.79%
Stixel level 88.25% 97.47% 87.80% 58.36%
Ground plane
estimation
Stixel
distance
estimation
Stixel
height
estimation
Stixel
semantic
segmentation
Stixel
visualisation
Methodology
LS
approximation
IRLS
approximation
Original
paper
This
paper
Line estimation V-Disparity estimation
MSc Computational and Software Techniques in Engineering 2016 (Digital Signal and image Processing Option)
V-Disparity
𝑳 𝒏𝒐𝒓𝒎 𝟏 𝑳 𝒏𝒐𝒓𝒎 2 𝑳 𝒏𝒐𝒓𝒎 𝟏 𝑳 𝒏𝒐𝒓𝒎 2