Digital Signal Processing Reference
In-Depth Information
Some tracking result frames are shown in Fig.3, which are used to illustrate the
effectiveness. Moreover, a quantitative comparison with the classical methods in based
on the total errors of center location in pixels is made, shown in Table II, in which the
numbers corresponding to the least errors are in bold. The experimental results above
demonstrate the effectiveness of the proposed online object tracking method.
Table 2.
Total errors of center locations in pixels
Mean Shift
[6]
Variance
Ratio [5]
Sequence
Proposed
Woman
180
84
16
Car
3430
276
92
Face
3805
4041
415
Fig. 3.
Several representative frames of 3 videos with bounding boxes located by proposed
method(red), simple template matching method(white), mean shift method(blue),variance ra-
tion method(cyan) and ground truth(yellow)
6
Conclusions
In this paper, we propose an stepwise Bayesian object tracking framework. For
sequential inference, we model the dynamical transition part based on particle filtering
in an affine sampling space, and the observation part via candidate likelihood
computation based on random KD-tree forest. The experiment demonstrates the
effectiveness of the proposed tracker in indoor and outdoor environments.
Visual tracking is a challenging issue due to the complexity of the imaging
environment. We plan to add template update scheme in the current framework so as
for occlusion and illumination variation handling. Moreover, for specific applications,
more efficient and robust observation model may need to be constructed to enhance
robustness of the proposed algorithm.
