We propose a technique to recognize actions of
grasshoppers based on spectral clustering. We track
the object in 3D and construct features using 3D object
movement in segments of video which discriminate between
different classes of actions.
We start by tracking the grasshopper using a
stereo camera setup. Then employ a background subtraction technique to track the object.We define a set of motion features based upon this
tracker output and perform spectral clustering on these W dimensional
feature vectors to classify object actions.
We tested our algorithm on 10 minute video of an individual grasshopper.
After clustering the actions of the grasshopper, we used the video synopsis technique of Rav-Acha et al. (CVPR 2006) to create short videos displaying each cluster of similar grasshopper actions. These videos allow one to see in which regions of the cage the grasshopper performed various actions.
Video synopsis results:
For more information, contact Maryam Moslemi Naeini