tld算法.ppt

Object Tracking with Tracking-Learning-Detection
******@
编辑ppt
Tracking-Learning-Detection
编辑ppt
retreal experts)
structure
P-expert: assumes that the object moves along a trajectory.
N-expert: assumes that the object can appear at a single location only.
编辑ppt
PN-learning
The car is tracked from frame to frame by a tracker. The tracker represents the P-expert that outputs positive training that due to occlusion of the object, the output of P-expert in time t+2 outputs incorrect positive example. N-expert identifies maximally confident patch (denoted by a red star*) and labels all other detections as negative.
编辑ppt
Tracking-Learning-Detection
P-N Learning
Object Model
Object Detection
Tracker
Integrator
Learning Component
编辑ppt
Object Model
Object model M is a data structure that represents the object and its surrounding observed so far.
Similarity between two patches is defined as
编辑ppt
Object Model
Nearest Neighbor (NN) classifier
A patch p is classified as positive if otherwise the patch is classified as negative.
A classification margin is defined as
编辑ppt
Object Model
Model update
(i) the patch’s label estimated by NN classifier is different from the label given by the P-N experts.
(ii) patches where the classification margin is smaller than .
编辑ppt
Tracking-Learning-Detection
P-N Learning
Object Model
Object Detection
Tracker
Integrator
Learning Component
编辑ppt
Object Detector
Scanning-window grid
scales step = , horizontal step = 10% of width, vertical step = 10% of height, minimal bounding box size = 20 pixels
Cascaded classifier
编辑ppt
Object Detector
Patch variance
This stage rejects all patches, for which gray-value variance is smaller than 50% of variance of the patch that was selected for tracking.
The stage exploits the fact that gray-value variance of a patch p can be expressed as ,
and that the expected value E(p)