The Sensor Systems Department of the Advanced Technology Centre has worked on many target tracking and data fusion systems over the last couple of decades.
One of the key tools used for this work is our Multsen simulation program.
Standard Kalman Filter
The Kalman filter and its variants form the basis of most target tracking systems. Its popularity stems from the fact that it is the optimum solution (for estimation of linear systems with linear measurement process and white Gaussian measurement errors) and that many of the real world tracking problems satisfy these constraints (at least approximately). Other key components of the standard algorithm are track initiation and measurement to track association and it is often these processes that determine the overall performance of the system.
Multiple Hypothesis Tracking
For dense multitarget environments or environments with large numbers of false detections more sophisticated tracking algorithms such as multiple hypothesis tracking are used. An ambiguous situation where two or more measurements might belong to a single track (or two or more tracks are competing for a single measurement) can be resolved by delaying the hard assignment of measurements to tracks until more data is available. This is achieved by creating a track branch (updated with a Kalman filter) for every plausible pairing of plot and track with the likelihood of each pairing being computed. Mutually compatible tracks are held together as hypotheses and the probability of the hypotheses are calculated. Subsequent measurements will result in some tracks increasing in likelihood whilst others decrease. Weak tracks and hypotheses are terminated when their probabilities fall below a threshold. This algorithm performs better than the standard Kalman filter in difficult environments but is more expensive in terms of computer resources.
Fusion of Attribute Data
Many sensors are capable of measuring some physical characteristics of targets; this may be size (real or apparent), image shape or perhaps a count of the number of engines on an aircraft. As with target tracking the aim is to identify which measurements belong to which objects and to combine these measurements into more accurate estimates of the target type. The Advanced Technology Centre has experience of using both Bayesian inference and Dempster-Shafer evidential reasoning. The Bayesian approach involves repeatedly applying Bayes’ equation to each new set of measurements. Each sensor is characterised by a set of a priori probabilities and each time a new measurement arises Bayes’ equation is used to calculate new (a posteriori) probabilities. Bayes is a special case of the more general Dempster-shafer approach.
Track Before Detect
In high false alarm environments the CFAR mechanism will reduce the probability of detection of real targets making conventional track initiation very difficult. Track before Detect overcomes these difficulties by processing the radar data before the conventional thresholding (i.e. detection) algorithm has been applied.
A constant range rate track in range time space (i.e. a straight line) will, under the Hough transform, become a point in Hough space (2 dimensional constant velocity motion is handled in a similar way). The Hough space is quantised into cells and those cells containing real targets will have many transformed points whilst the false alarm data will be spread over many cells (since straight lines formed through the false alarms will be less consistent). For those cells exceeding a threshold the inverse transform is applied to extract the track parameters. The technique has been demonstrated with simulated data and has also succeeded in detecting targets from recorded radar data.
