1. Introduction: Null Space View Invariant Representation for Motion Event Recognition, Classification and Retrieval

A fundamental set of 2-D affine invariants for an ordered set of n points (not colinear) in two dimensional space is expressed as an n-3 dimensional

subspace. The effectiveness of our null space invariant (NSI) representation for motion event classification and retrieval is due to two factors:

(1) This particular space is invariant to any linear transformations, and therefore greatly improves the accuracy of classification and retrieval for motion events

from different points of view due to camera motions.

(2) It does not need any assumptations and preserves all the information of the original raw data.

Overall, the NSI based motion event classification and retrieval method is much more accurate than traditional view invariant classification and retrieval methods such as

the approach of Bashir et al. ACM Multimedia Systems Journal relying on curvature scale space (CSS) and centriod distance function (CDF),

Bashir et al. TIP based on hidden Markov models.


The approach is described in:

Xu Chen, Dan Schonfeld and Ashfaq Khokhar

Robust Null Space Representation and Sampling for View-Invariant Motion Trajectory Analysis

CVPR 2008 [pdf]

If you use this code in your research and publication, please refer to this paper.

2. Software

In the following, we display the demo for our method. In the demo, we use 40 classes for view invariant motion event classification and retrieval.

The demo is implemented with Matlab. For any further questions, please contact Xu Chen with xchen27@uic.edu