Digital Signal Processing Reference
In-Depth Information
Independent Tracking Without Fusion
Independent tracking without fusion is to track the regions of object in multiview
video by segmentation of each frame in individual view.
Cigla et al. [
3
] presented a multiview video object segmentation algorithm by
integrating color, depth, and motion features. A region-based color segmentation
algorithm based on modified Normalized Cuts is firstly adopted to generate over-
segmented segments. Depth map is then estimated for subregions in the available
segmentation mask by region-wise planarity assumption. Multiview video segmen-
tation is extended from image segmentation by combining the color and depth with
additional optical flow information to provide the motion field.
Independent Tracking with Fusion
Independent tracking with fusion is to segment the tracks in each camera stream
and then project the tracks to another camera view or a common view (ground plane
[
27
,
49
], “plan-view” [
40
]), or collect the 2D local tracks from individual view to a
global 3D track [
29
] or central node [
43
].
A multiview segmentation and tracking system in cluttered scene with mul-
tiple people is presented in [
27
], which is named M2Tracker. Exploiting the
approximate object's shape and location prior helps the segmentation of each
view using Bayesian classifications. The region-based stereo algorithm is capable
of finding the 3D points inside the object. By combing evidences from different
camera pair and producing feet-region likelihood estimation on the ground plane,
globally optimum detection and tracking of object is attainable using Kalman filter.
Zhao et al. [
49
] presented a similar and realtime system that detects and tracks
object independently for each stereo camera, and integrate tracking results from all
camera pairs to a multi-camera tracker (McTracker), which track each object on the
ground plan. An object tracking framework based on dynamic Bayesian formulation
is reported in [
40
] to observe and track object on the plan-view map by combining
local appearance feature and stereo depth data.
Instead of projecting the multiple single-view tracks to a common view, com-
bining the tracked 2D object into a 3D tracking module is another strategy for
multiview data fusion. In [
29
], following the people detection using background
subtraction and human-template correlation, 2D objects are tracked separately in
each of camera by a graph matching. A 3D tracker is established using geometrical
consistency between 2D objects to estimate the 3D head position. For tracking large
numbers of tightly-spaced and rapid-moving objects, i.e., hundreds of flying bats, a
multiobject multi-camera tracking framework is proposed in [
43
]. It maintains the
sensor-level tracking in each view and single-view measurements send to a central
node for across-view data association and tracker fusion. The feedback from central
node is then used for adjusting sensor track with across-frame data association.
