“Visual tracking is the problem of estimation of a time - variant metric parameter (the kinetic stateof the target) from uncertain observations of the parameter.” – Bobby Rao “Data Association Methods for Tracking Systems”
Visual tracking is divided
into two important parts.
The most popular technique foroptimal estimation is the recursive linear estimator: Kalman filter, which haswide applications in visual tracking. The structure of this Kalman filter isvery simple. The following figure shows the structure of the Kalman filter.
|
|
|
|
|
|
Data association algorithms
areclassified into two groups:
The optimal approach is calledMultiple Hypothesis Data Association. This in turn involves all combination ofmatches between previous and current observations, which are then calculatedfrom time to time. At each time step a hypothesis tree is grown, each branchrepresenting a particular combination. Unfortunately this cannot be implementeddue to the exponential increase in the number of branches as the number ofobservation points increases.
The suboptimal approach isfurther divided into two popular methods known as Nearest Neighbour data associationand All Neighbour Data Association.
The Nearest Neighbour methodfails in the case of high clutter (additional unwanted measurements) density,when there is more than one target and when the tracks intersect giving way toa major disadvantage of order dependency in multi target cases. Regardless ofthese drawbacks this method is computationally inexpensive, works well in lowclutter density, and in the case of non-interfering multi targets. In additionto this the track management is also simple.
The All Neighbour approachsolves the problem of order dependency but is very expensive computationally.This approach involves two main implementations known as Probabilistic DataAssociation filter (PDAF) and Joint probabilistic Data Association filter (JPDAF).The PDAF gives a kind of an average target taking into consideration all thevalidated data points and updates the filter. This method also suffers seriousdrawbacks when the validation (the process of removing falsely generatedobservations) regions overlap or when the regions increase in size. The JPDAFis very expensive and there is no track initiation or track removal in PDAF andJPDAF.
Due to many drawbacks in the caseof cluttered environment, Donald B. Reid in his paper “An Algorithmfor Tracking Multiple Targets” developed an algorithm to track multipletargets in cluttered regions, which is capable of initiating tracks, accountingfor false or missing targets or reports and processing sets of dependentreports. A Kalman filter is used for the update of the target states resultingfrom each data association hypothesis. The development of this algorithm isvery useful as multi target tracking has wide range of applications in bothmilitary and civilian areas. For instance applications include detecting enemyaircraft, ballistic missile defense, air control traffic etc. This algorithmtakes measurements from two different types of generic sensors, Type 1 and Type2 sensors. The first type is capable of giving the information about the numberof targets in the surveillance region unlike the Type 2 sensor. Type 1 sensoralso processes the data in the form of batch processing whereas the second typeprocesses one data set at a time. The examples of the first type are a cameraand RADAR. An example for the second type of sensor is a RADAR detector.
Data Fusion is also a majorfactor in visual tracking. A target is tracked based on many features likecolor, shape, texture, motion, position etc. each of these features are rankedand the best are combined or fused to get a better tracking. In many cases onlyone feature is considered. But many people have come up with different datafusion and ranking techniques, which combine more than one cue or feature inorder to get a result that is very close to the target and it’s originalfeatures.