Johns Hopkins Computer Science


Robotics related projects in the CIRL Lab

Relevant publications

Vision-Based Navigation

Floor Suppression

This model and values from the disparity map of a stereo camera setup facilitate the computation of two essential parameters: the camera's height and the camera's tilt angle. By choosing rows of the disparity map that are assumed to be floor, for example, two rows towards the bottom of the disparity map, two histograms and their corresponding highest frequency disparities. These two floor disparities can be applied to the model, and the aforementioned camera characteristics will be the result.

With the camera height and tilt angle known, expected floor disparities can be guessed at any coordinate in the disparity map. Consequently, scanning through the disparity map, if an estimated disparity matches the actual value in the disparity map, that pixel can be considered floor. Iterating through while repeatedly zeroing out any matched floor pixels will result in a disparity map with the floor suppressed.

Paper Visible (above floor)
Paper Suppressed (on floor)

Obstacle Avoidance (MPEG Movie)

Once the floor has been suppressed, the remaining non-black blobs represent potential obstacles.


Assuming the floor has been suppressed and a region of interest has been selected, the region of interest can be continuously tracked by assuming the middle pixel of the region of interest lies on the object being tracked. The middle pixel also gives an estimation to the disparities of the object. A histogram can be computed of disparities that lie within a range around the middle pixel's disparity. The most frequent disparity represents a peak in the histogram. In the histogram, the width of the peak specifies the varying disparities that the camera has computed for the object. Within this padded region of interest, the largest blob with pixel disparities that lie within the computed range indicates the objects new position.

For subsequent disparity maps acquired from the frame grabber, the floor can be suppressed, locally around a padded region of interest, using the same method described earlier. Tracking can then proceed as mentioned previously.

For more information:

Gregory D. Hager ( Faculty
Jason J Corso ( Webmaster
Linda Rorke ( is our departmental contact.