please discuss Real-Time Object Detection for smart vehicles here.

Edited by author 03-21-2006 10:56 PM
The paper by Gavrila discusses a (near) real-time framework for object detection. On some level matching using distance transforms has been a proven technique for shape matching. The chamfer distance based matching technique has been used by the authors of the paper, however, I presume similar results could be obtained using a generalized hausdorff distance metric. I would be interested in seeing if there are any advantages of using one over the other? The authors report that using chamfer results in speed up but I guess we can also do hausdorff distance computation in linear time.

Given that the technique is based on the performance of the underlying contour segmentation the results for pedestrain detection seem to be good. However, the method does not seem to work for considerable scaling of objects that exist in the usual perdestrain like scenario. A possible extension as suggested in the paper would be to use stereo for foreground / background distinction and IR imaging for pedestrian silhouettes.

I wonder how this method would fare if included in our viola jones vs schneiderman comparisons on pedestrian detection.

Might a Gaussian blur work just as well as the other DT distance measures they discuss? I like the DT idea but the parameters are a bit ad hoc. It seems like there should be a way to choose an optimum distance measure based somehow on your template and search strategy.

A few parts of this paper seem a bit vague, and I'd like a little more detail about what's going on. I'm also concerned with their false-positive rate. In both test cases, they mention one or two false positives per image. In the case of matching circle and triangle road signs, they take care of the problem using a mysterious RBF network, but nothing is said in the case of pedestrian detection.