Target Classification With Radar and EO Fusion on USV Platforms

Modern electro-optical (EO) systems on USVs provide a basis for target recognition using high-resolution daylight and thermal cameras combined with intelligent processing. Recent developments in machine learning have significantly improved classification capability, particularly where large training data sets are available.

Wide-angle cameras can detect targets of interest, but radar provides a complementary and generally more resilient method. It supports detection at longer ranges and in low-visibility conditions. In a typical USV-based deployment, a maritime radar detects and tracks targets before passing information to a controller that engages the EO system. The camera can then zoom onto the target to confirm whether it is relevant and identify it by comparing the image with a trained database.

Radar Acquisition and Classification

Automatic track acquisition from radar is a well-established process. Radar video is processed to identify target-like detections, known as plots. The plots are then correlated over time to identify consistent behaviour that indicates a target rather than clutter. Tracks may be maintained for a configurable period to build up information about the target's dynamics, including its speed, course and consistency of motion. Longer acquisition times provide greater confidence and better estimates of speed and course, but take longer to establish the track.

Information about a target's position and motion can be combined with its radar cross-section, which indicates target size, to provide a first level of classification. Classification can be based on the likely target type, using the observed radar cross-section, image and dynamics. It can also reflect the perceived threat, based on the target's position, speed and direction of motion. For example, a target on a course that intersects with own-ship may be considered a risk or threat.

Within Cambridge Pixel's SPx Server radar tracking system, radar-derived targets are classified using a rules-based approach. Details of each target are passed to a configurable set of rules, which combine properties and conditions to assign the target to a class. The rules take into account the target's position, speed, size, Doppler, age and course, including whether its course indicates a collision approach to own-ship. Information about known static features can also be incorporated into the rules, allowing returns from navigation buoys, clutter, or landmass to be eliminated.

Camera Control and Classification

Newly acquired radar tracks can be used to slew a camera into position for visual review of the target. Radar information such as range, bearing and size can be used to position the camera appropriately. The camera can then be stabilised on the target by compensating for own-ship heading and, ideally, roll and pitch. The view can be dynamically repositioned using updates from the radar system (slew-to-cue), and/or the camera system can provide its own control using analysis of the image data (video tracking). In either case, the camera-derived imagery can provide a basis for accurate bearing location and target recognition.

Modern AI-based processing can recognise targets using EO data from daylight and thermal imagery. This depends on a good training set of representative target types. The image data can then be compared with the trained data set to find the closest match. Classification is typically into a broad object class, such as vessel, motorboat, sailing boat, buoy or person in water.

Fusion of EO and Radar Tracks

Target information from the EO system can be fed back into the fusion system to add detail to the target record. If the radar and EO tracks are derived independently, the association process must match the two sets of target reports and confirm that they relate to the same real-world target. If the EO-derived track is based on information from a radar track, the track reports are implicitly associated and the fusion process simply aggregates the information.

Combining radar-derived and camera-derived information allows the fusion system to make decisions about the nature of the threat.