MOVING TARGET DETECTION WITH COMPRESSED SENSING
The detection of moving targets is one of the central tasks of radar. It is possible, however, that the number of measurements available is inadequate to form the basis for clear and sufficiently accurate statements. Additional know-how from other sources can be helpful here. Compressed sensing is one example.
Information on movements on the ground has, for many decades already, been extracted from airborne radar measurements (Ground Moving Target Indication, GMTI). These offer the possibility to detect military or illegal activities, irrespective of the time of day and cloud cover. The observed scene is, however, dominated by contributions from stationary objects which are superimposed on the measurement data of the moving objects. Modern GMTI methods such as STAP (Space Time Adaptive Processing) use short radar pulse sequences, with the result that speed and position changes of the target no longer have an influence. On the other hand, STAP requires several receive channels arranged side-by-side as well as additional comparative data from areas with no moving targets.
One active field of research in this context is the utilization of external information not only to improve detection but also to replace non-measured data. The latter, although only possible to a limited extent, is necessary as suitable measurements cannot be acquired at all points in time. One highly significant piece of information, although seemingly trivial at first glance, is the observation that pronounced man-made movement only occurs in a few – previously unknown – locations on Earth while the largest section of the scene is only moved, at most, by the wind.
This sparse distribution forms the basis for compressed sensing (CS). Although this relatively new mathematical theory is constantly being refined and further developed, areas in which it offers real added value have already been identified. One of these is GMTI. Fraunhofer FHR recognized the potential of CS at a very early stage. Accordingly, the institute co-initiated and organized the "International Workshop on Compressed Sensing Theory and its Applications to Radar, Sonar, and Remote Sensing" (CoSeRa) which has now taken place for the fourth time.
One of the known disadvantages of CS is the longer computing time that is required compared to traditional techniques. This is due to the fact that further processing only takes place once the scene has been reconstructed from the measurement data. Traditional techniques, on the other hand, filter the data to remove everything except the moving targets. Although modern computer systems have adequate computing capacity, it is still more advisable to break the measurement data down into smaller chunks and process each of these individually. It therefore proved effective – analog to the traditional STAP approach – to observe short radar pulse sequences in several receive channels and divide these based on distance to the radar. CS can then use the data to reconstruct a diagram as in Fig. 1, which presents the scattering intensity of a corresponding object on the ground for each line of vision and Doppler frequency. Here, the Doppler frequency is proportionate to the change in distance between the observed object and the radar.
The black mark in Fig. 1 highlights a point-shaped moving target. As the radar itself is moving, each stationary place on the ground also has a relative speed depending on the angle of vision. As a result, the stationary ground in Fig. 1 is shown with a red band. The original CS approach is, however, only designed for random distributions and, for this reason, a number of modifications were successfully tested at Fraunhofer FHR. These modifications allow a stable reconstruction, in spite of the band structure.
With knowledge of various measurement parameters, it is possible to calculate where the ground contributions were reconstructed in Fig. 1. Accordingly, a good distinction can already be made between the ground and moving target section, without having to use statistical methods to find the band in the reconstruction. All moving target detections from all distances and radar positions are shown on a map in Fig. 2. A convoy of eight vehicles moving along a road delineated in blue can easily be recognized. The number of false alarms outside the convoy is minimal.
On the other hand, the ground sections of the reconstructions can also be gathered and used to create a SAR image, as shown in Fig. 3. This is free of moving objects which would otherwise be superimposed over the entire image thus potentially masking interesting information. The noise level in the image was also reduced through CS processing to facilitate further analysis.