The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
612 
single target tracking it only need to depend on one feature 
property, but in multi-target tracking it may need a integration 
of different kinds of features for directing at proper target, and 
it also could using some suitable ways, such as filter methods 
for multi-target. 
4.1 Object Modeling 
Object modelling is a representation of object, in other words it 
utilizes one feature characteristic or the combination of features 
to express the object. The object’s feature could be contour, 
shape, color, position, texture, velocity and so forth. The more 
features included, the easier to identify the object. But the 
combining features will increase burden of processing and 
demand composite methods. To construct the model of object, 
we can use the features directly, or transform them into other 
forms such as templates. 
Features of the object may change during the course of 
tracking, so it requires that the model should be adaptive to the 
changing or other influences, for example occlusion and 
unexpected movement. This is considered as the robustness of 
model. There are many ways to make the model more stable, 
including using multi-features model and updating the model 
over time. 
4.2 Object Tracking 
Using prior information that forms the model of object, tracker 
predicts the object’s position in succedent frames. 
Corresponding to different models, object tracking has different 
methods. Object tracking methods attempt to ascertain the 
coherent relations of feature information between frames, and 
the strategy of it is no more than searching and matching. 
Hausdorff distance is a valid measurement for shape and texture 
features of the object. It can create sparse point sets with feature 
detectors in images, and the point set of image region labelled 
as the object is the object’s model for Hausdorff measurement. 
It is able to tackle the deformation of object, because it 
describes the contour and texture of the object with bulk of 
points. Taking the measurement and the model, it translates 
object locating into the matching of point sets (Huttenlocher et 
al„ 1993). 
Motion is a kind of state. A typical motion state vector is 
composed of the object’s position, velocity and acceleration 
along each direction. If the prior and current states are known, 
the posterior state will be predicted. It is feasible to resolve the 
problem of object tracking by state estimation means. Kalman 
filter is one of the state space methods. To define it, the 
Kalman filter is a batch of mathematic equations that solves the 
least-squares question recursively. It predicts the values of 
current state utilizing the estimation values of former state and 
the observation values of current state, executing the procedure 
recurrently until the values of every state estimated. To get the 
estimation values of each state, all the previous observation 
values have been involved. For object tracking, the state 
equation is the model of object in Kalman filter, and it describes 
the transfer of states. The observation is the position of object, 
and the state vector like mentioned above contains position, 
velocity and acceleration. Putting the positions of object 
detected in initial frames into the observation equation of 
Kalman filter and taking the accurate positions as the initial 
value of state variant, it compares the output of filtering with 
precise result to testify the correctness of initial input. It repeats 
the process until the filter is stable (Forsyth et al., 2003). 
Mean-shift algorithm is an approach that searches the maximum 
of probability density along its gradient direction, as well as an 
effective method of statistical iteration. Object tracking with 
Mean-shift algorithm is another class of technique that locates 
the target by modeling and matching it. Both the modeling and 
matching are performed in a feature space such as color space 
and scale space. The mode of it is using the relevant similarity 
measurement to search the best match. The object tracking 
basing on Mean-shift algorithm mainly processes on the color 
feature. Choosing an image region as the reference object 
model, it will quantize the color feature space, and the bins of 
the quantized space represent the classes of color feature. Each 
pixel of the model can corresponds to a class and a bin in the 
space, and the model can be described by its probability density 
function in the feature space. Instead of PDF (probability 
density function), it takes the kernel function as the similarity 
function to conquer the lost of spatial information. Another 
reason for using kernel function is smoothing the similarity 
measurement to ensure the iteration converge to the optimized 
solution during search (Comaniciu et al., 2003). An object 
tracking result of airborne video using Mean-shift method is 
shown in Figure 6. 
Figure 6. An object tracking result of airborne video using 
Mean-shift method 
5. SYSTEM FRAMEWORK 
To the technical approaches analysed above, it needs a 
framework to integrate all these methods. For the technique of 
moving target detection and tracking divided into three parts, 
each part would be an isolated module for its independent 
function in applicable system. Therefore, the processing is in 
and between different modules. There are many systems 
employ a series procedure. Compensation comes first, the next 
is detection, and tracking put on the last. The reason of that is 
anterior module always be taken as the precondition of 
posterior module, and results of each one could be inputs of the 
next one. However, this kind of system is not considering the 
interactions between different modules. For example, the result 
of segmentation can be the initial value of compensation, and 
the tracking result can accelerate the detection processing. 
As shown in the figure 7, distinguishing from traditional 
technique framework, the presented system framework 
introduces two more modules, which are data capture and 
collaboration control. Data capture module gets the video image 
data and samples it into image sequence, and then it will 
distribute them to another three modules that are the central 
parts of the system. The three modules implement a parallel 
processing, and this will lower the cost of time. After the 
interior computing, they transfer the outputs that always in the 
manner of parameters to collaboration control module. The 
control module manages all the other modules by sending 
orders to them, and it provides interface to user and exterior 
system.