Human Action RecognitionWeek 2

Taylor Rassmann

Action Recognition from One Example

Action is often distinguished from activity in the sense that action is an individual atomic unit of activity. In particular, human action refers to physical body motion.

Action Recognition from One Example

Novel feature representation that is derived from space-time local (steering) regression kernels (3D LSKs)

This feature representation can attain data even in instances of high distortion and data uncertainty

Action Recognition from One Example

This is achieved by measuring the likeness of a voxel to its surroundings based on computation of a distance between points.

These points are measured (along the shortest path) on a manifold defined by the embedding of the video data in 4D

For better classification performance, space time saliency detection is applied to larger videos to crop to a shorter action clip

Action Recognition from One Example

The key idea behind 3D LSKs is to robustly obtain local space-time geometric structures by analyzing the photometric (voxel value) differences based on estimated space-time gradients, and use this structure information to determine the shape and size of a canonical kernel (descriptor).

Approach Taken

Use of pair wise distances of salient regions

Saliency extraction complete from a few actions in KTH data set

Current Work

Code for saliency statistics in progress This will implement the distance metrics

for pair wise distances between features Possible use of thresholding for some of

the salient regions to prevent merging This will help acquire different parts of

features Note: Thresholding must be careful not to be

to high or low to eliminate some of the important data