Schematic overview of my research world regarding free-space segmentation using stixel algorithms.

 

Illustration of extended stixel world framework  

 

Step 1 (related paper: ITSC14)

Fusing color with disparity in an Extended Stixel World to increase robustness against adverse imaging conditions.

 

Step 2  (related papers: ITSC15 & PPNIV15)

Taking disparity analysis out of the critical path without degrading segmentation quality to reduce system response time/latency.

Illustration of the FCN framework

   

Step 3 (related paper: arXiv16 & EI-AVM17; view corresponding poster)

Using the same strategy of online, self-supervised training as in step 2, but now rely on the rich encoding of Convolutional Neural Nets for a more advanced and specialized color modeling. We use a small net that is initialized with offline training and tuned online so that we achieve fast convergence. Additionally, we achieve fast inference by executing it as a Fully Convolutional Net, using the CN24 library.