Презентация на тему «Scene-Consistent Detection of Feature Points in Video Sequences»

Scene-Consistent Detection of Feature Points in Video Sequences

Ariel Tankus & Yehezkel Yeshurun

CVPR - Dec. 2001

Tel-Aviv

University

Outline:

Relating convexity-based detection of feature points to scene geometry. Feature points tracking algorithm. Comparison with two other methods. Measures for evaluation of tracking algorithms w.r.t 3D scene-consistency.

Task Definition:

Robust detection of scene-consistent features in video sequences.

Object recognition. Correspondence points for recovering 3D characteristics of the scene.

Convexity.

Goals:

Intrinsic Property:

Operator for Feature Detection

?

Detect convex or concave image domains.

Detect local “circles” where the gradient of the intensity function points outward along the whole circle.

The gradient points in all orientations along the “circles”.

(motivation)

Equivalently:

Operator for Extracting Certain Gradient Orientations

At the discontinuity ray of the arctan: Yarg??.

Darg - An isotropic variant of Yarg.

Response of Yarg to the Intensity Surface

Examine Yarg in well behaving image domains. Intensity is twice continuously differentiable.

We examine all possible intensity configurations. Four of them lead to infinite Yarg response.

The basic observation:

Response of Yarg to the Intensity Surface (cont

?

The cases include: Some configurations where is a local extrema of , and some configurations where one side of is flat, but the other is convex or concave.

Only specific differential geometry structures of the intensity function causes Yarg??.

Response to Local 3D Scene Structure

?

?

Yarg responds to certain geometric features of the 3D scene object.

Yarg?? for certain elliptic, hyperbolic or parabolic points on a Lambertian 3D surface illuminated by a point light source at infinity.

Tracking Algorithm

Stable points: points where . These points are the only input to the point tracker.

toys

#28

#8

#16

#36

#24

#44

parking

#200

#50

#100

#225

#150

#250

traffic

#5

#55

#65

#10

#70

#15

Evaluating the Performance of the Algorithm

Two measures for evaluating performance of scene-consistent point tracking algorithms. Each measure aimed at a different task: Maximal tracking time. Correspondence of points in successive frames. Their common goal: to quantify the consistency of tracks with 3D scene.

Measures for Evaluation of Scene-Consistency

Completeness: A track is complete if the same 3D scene point is being tracked, up to a certain level of noise, in every frame where it appears.

Correct 3D point of track T = 3D point tracked for the longest time under track T

Measures for Evaluation of Scene-Consistency (cont

Stability:

Tracking Comparison

We compare the Darg-based algorithm with two other algorithms: Junction detection (Lindeberg). with automatic scale selection. Tracking by Kalman filter. KLT (Kanade-Lucas-Tomasi). Tracker based on affine image change model. Features maximize tracking quality.

Completeness:

Stability:

Experimental Results

Darg is more stable the Junction detection, and sometimes more than KLT. Sometimes Darg equates with KLT. Darg completeness is at least comparable to that of Junction detection or KLT, and sometimes even better. Darg has significantly lower no-tracking time (Darg: 4, KLT: 81, J.D.: 121 frames).

Summary

Convexity-based method for scene-consistent feature points detection in video sequences. Detection relates to specific features of the intensity surface. These intensity features relate to geometric features of the 3D object.

Summary (cont

A stable point tracking algorithm is described (2D Kalman filter). Two measures serve in a comparison with two other tracking methods. Completeness: Maximizes tracking time of a 3D scene point. Stability: Consistent tracking of 3D points between successive frames.