Danvil's Blog

I am David Weikersdorfer a PhD student from Munich/Germany. This is my personal blog where I will post about my hobby game programming and about my research in computer vision and robotics.

[danvil-blog]

Event-based SLAM for embedded robotics

Perhaps you remember my post some month ago about event-based particle filtering? In the meantime we extended the particle filter to a full SLAM algorithm. SLAM is the abbreviation for "Simultaneous Localization and Mapping" and refers to the fundamental problem in robotics and computer vision to generate a map of an unknown environment during navigation. Your intelligent vacuum cleaner or bring-me-beer-robot will use SLAM :)

Our approach to SLAM is highly efficient and my ad-hoc implementation runs already in 10x realtime. This makes our algorithm especially suitable for embedded systems.

I have been to Sankt-Petersburg in Russia last week and presented the event-based SLAM method at the conference ICVS 2013. Sankt-Petersburg is a crazy city. They are palaces and gold everywhere while at the same time most of the city is decaying...

[danvil-blog]

Ludum Dare 26: Minimal (Art)

I complete forgot to write about ludum dare 26. The theme was minimal and I tried a grenade throwing game with minimal art.

And of course Piet Mondrian had something to do with this. I tried to build a Mondrian-Generateor. Actually I am quite proude of it :)

As usual you can play my entry here and the best entries are listed here.

[danvil-blog]

Ludum Dare 25: "Oppression"

It's time for Ludum Dare 25. Sadly, this time I could only start working Satuarday afternoon because I was in China on a robotics conference. So jet-lagged and generally distracted, I settled with the classic police-beats-humble-citizens theme... In retrospective the choice is interesting, but China wasn't that bad - actually not bad at all.

Due to my positive experiences at the last Ludum Dare, I worked again with Unity 3D. My goal this time was to simulate lots of small people running around and I could use similar code as in my Evolution entry. However the game is more unfinished than usual due to the lack of time. Have a look at my entry or play it directly.

[danvil-blog]

Event-based Particle Filtering for Robot Self-Localization

For those more interested in science, a small remark about my recent scientific research. I published an interesting, new method for robot self-localization using a low-profile event-based sensor. See here and here for more information about the sensor.

The main challenge was to adapt well known methods in computer vision to a new kind of hardware. Instead of full images at a fixed frame rate our method works with a stream of individual pixel events. This has the advantage that much less computation power, almost no memory and much less electrical power is required - making our method especially suitable for embedded applications.

Next week I will fly to China and present the paper a the Conference on Robotics and Biomimetics - hooray! This will be an exciting trip and I am really curious what China is alike.

[danvil-blog]

Ludum Dare 24: "Evolution hates you"

Finally! Evolution is the theme of Ludum Dare 24. In case you do not know, Ludum Dare is 48 hours game programming contest. Evolution has been one of my favourite topics for a long time (just have a look at these ants) and I am a bit excited to program a game about it.

This time I decided to finally give Unity3D a try and it actually worked out quite nicely. My game has some small procedural generated critters which mimic their behaviour in their appearance. Have a look at my entry "Evolution hates you" or play it directly using Unity Webplayer if you are under Windows.

[danvil-blog]

Depth-Adaptive Superpixels

I will present a paper about "Depth-Adaptive Superpixels (DASP)" on the 21-st International Conference of Pattern Recognition (ICPR 2012).

Depth-adaptive superpixels is a novel oversegmentation algorithm for RGB-D images. RGBD images have become incrasingly popular due to the Microsoft Kinect.

In contrast to previous approaches, DASP uses 3D information in addition to color. DASP can be used as a preprocessing step for various computer vision applications, for example image segmentation or point cloud segmentation.

The paper, the source code and a small dataset can be found on the github project page Danvil/dasp.

[danvil-blog]

Source Code for Darwin Voxel Engine

Today I published the source code for the Darwin Voxel Engine on GitHub. Have a look at the public repository! There are two examples to try out: the Cornell box and the indoor scenario. The landscape scenario is coming in the near future.

[danvil-blog]

Ludum Dare 23

Last weekend I participated in the Ludum Dare 48h game contest. It was the 10th anniversary and Ludum Dare got over 1000 entries!

My game is a small rogue-like where you have to find good equipment and most importantly rescue kittens! The programming was done with Java using Netbeans IDE. As a game engine I used slick a minimalistic 2D game engine. Pixel art was created with GraphicsGale and Paint.NET. Sadly I did not have enough time to try out sound and music. In the end I also had to skip some other features like a proper intro and rewarding game mechanics. This was the third time I participated and the first time I completed something :)

Go and have a look at all the other games! There are some real gems among them.

[danvil-blog]

Indoor scenario with the Darwin voxel engine

Being sick at home, I started an indoor test room using the Darwin voxel engine. The room is made of 250000 voxels and lighting is computed using the radiosity lighting algorithm. You can walk around in it and delete voxels in realtime.

[danvil-blog]

Voxel Engine Screenshots

"Darwin" is the working name of a game engine I am developing in my free time since about a year. The engine uses voxels and thus looks a bit like Minecraft, but I want to focus on new technology which you will (probably) not see in Minecraft anytime soon.

The engine is developed in C++ and uses OpenGL for rendering. I am only using basic libraries (boost, freeglut, GLEW and DevIL) - all engine code is written by myself with help from a friend. Is runs under Ubuntu 11.04 and Windows 7 on a normal desktop computer and my laptop.

To get started some screenshots:

~56 million cubes (13824 cells of 16x16x16 cubes each) with perlin noise based terrain generation

Cornell box with cubes (side length 144 cubes) - global illumination lighting for cube sides is computed using Monte-Carlo ray tracing

Cornell box with cubes (side length 64 cubes) - lighting for cube sides is computed using the radiosity global illumination algorithm

About 15 million cubes shaped to form a big sphere

[danvil-blog]

This is the latest application of my markerless human motion tracker. You can find more about my research at my homepage at Technische Universitaet Muenchen.