Moving/Static Classification via Tracking 


This project deals with the problem of classifying dynamic objects out of the entire objects detected in sequential frames taken from a single moving camera and tracking them along the frames.

The challenge derives from the difficulty to seclude the dynamic objects from the static ones in a situation where all objects seem dynamic because of the motion of the camera,

In addition, the conditions of the inputs can vary a lot from each other (homogeneous background, urban background, multiplicity of dynamic objects, different speeds of camera motion, angle of filming, quality of the frames taken…) .

Algorithm modules

1.Interest point detection on the input images.
2..A creation of a graph of K-NN (K nearest neighbors ) with a constraint of maximal distance between neighbors.
3.score every edge in the graph by using the homogeneous scale assumption.
4.Groups segmentation using the graph data.
5.Unify groups which describe the same object in the image


frame 1 scoring:

frame 6 scoring:

frame 2 clusters:

frame 6 clusters: