3.3.2 Group Tracker

Tracking algorithm is implemented as a library. Application task creates an algorithm instance with configuration parameters that describe sensor, scenery, and behavior of radar targets. Algorithm is called once per frame from Application Task context. It is possible to create multiple instances of group tracker.

Figure 11 explains the steps the algorithm goes through during each frame call. Algorithm inputs measurement data in Polar coordinates (range, angle, Doppler), and tracks objects in Cartesian space. Therefore we use Extended Kalman Filter (EKF) process.

Point cloud input is first tagged based on scene boundaries. Some points may be tagged as “outside the boundaries”, and will be ignored in association and allocation processes.

Predict function estimates tracking group centroid for time n based on state and process covariance matrices estimated at time n-1. We compute a-priori state and error covariance estimations for each trackable object. At this step we also compute measurement vector estimations.

Association function allows each tracking unit to indicate whether each measurement point is “close enough” (gating), and if it is, to provide the bidding value (scoring). Point is assigned to a highest bidder.

Un-assigned points are going through an Allocate function. During the Allocation process, points are first joined into a sets based on their proximity in measurement coordinates. Each set becomes a candidate for allocation decision. It has to pass multiple tests to become a new track. Once passed, the new tracking unit is allocated.

During Update step, tracks are updated based on the set of associated points. We compute the innovation, Kalman gain, and a-posteriori state vector and error covariance. In addition to classic EKF, the error covariance calculation includes group dispersion in measurement noise covariance matrix.

The Report function queries each tracking unit and produces the algorithm output.