Research

 

Home Publications Research Projects Source Code Contact

 

Computer Vision

Traffic Sign and Pedestrian Recognition in an Urban Scenario

This work was done at Daimler Benz, Ulm, Germany. We have an E-class Mercedes equipped with a camera, an LCD monitor and a dual Pentium II PC. Here are some results:

day.avi (15.5 MB) - Traffic sign recognition during daytime. Note that all traffic signs that are detected properly but not classifed were not included in the classifier training set. This shows that the algorithm does not falsely classify closely resembling signs as one of the signs in the training set, which is exactly the behavior we want.
night.avi (7.27 MB) - Traffic sign recognition at night.
ped.avi (8.7 MB) - Pedestrian detection under different conditions.

Automatic Target Recognition

 

The slide show above shows some results obtained with the algorithm. The first image is the original range image, the second one is the SNF filtered image, the third one is the edge image obtained by filtering based on the heights above the ground on either side of each edge pixel followed by thinning, the fourth one is the edge image with noise edge points removed, and the final image shows the target center after the novel two stage hough transform. The pose of the target is also estimated accurately by the algorithm although it is not shown here. Here's a movie of the range data color coded for heights above the ground. As part of this project, i also had to implement some basic image processing algorithms. Here are movies showing the results of two of those algorithms:

Symmetric Nearest Neighbour Filter (SNN 5x5)
Thinning Animation