From Radar Returns to Safe Decisions: Radar Processing for Maritime Autonomy
Without a human onboard with all-around situational awareness, the sensor systems of autonomous and remotely operated vessels are critical for navigational safety and mission effectiveness. Much attention goes to the onboard camera suite, but radars, and their associated processing, are also a part of the sensor system mix.When developers specify radar for an autonomous vessel, attention often goes to factors such as range, resolution, antenna size, and technology. Those choices matter, but they do not determine what the autonomy system or remote operator ultimately receives. This is determined by the quality of the radar processing.An autonomy control system needs machine-readable information; a remote operator needs the maritime picture presented as clearly as possible to aid decision-making and reduce cognitive load. In both cases, a suite of sensors, with data processed and fused into usable information, is ideal.Why radar remains importantAn autonomous vessel may carry cameras, an Automatic Identification System (AIS) receiver, Global Navigation Satellite System (GNSS) receivers, LiDAR, sonar, and other sensors. Cameras can help identify an object, while AIS can provide a transmitting vessel's declared identity, position, course, and speed. However, neither alone provides a complete picture.Cameras can be affected by darkness, glare, fog, rain, and spray. Most can only provide contact bearing, or, at best, a rough estimate of range. Small and fixed contacts are unlikely to carry AIS transponders, and those that are fitted with AIS can have it switched off or incorrectly configured. However, radar can operate at long range in poor visibility and provide accurate position information for small and non-co-operative contacts.Radar is not infallible. Performance is affected by the installation, sea state, weather, target size and reflectivity, and processor settings. Radar should complement other sensors rather than replace them, but it remains an important independent source of information that autonomous systems should not ignore.From echoes to detectionsA radar will output video, in either analogue or, increasingly, digital form. In radar terminology, "video" means the stream of return intensities from the radar, as opposed to image-like camera footage. It is a constantly changing field of reflections, in polar space.The radar video picture may be cluttered and complex. A steel ship may produce a strong radar return, but so may a harbor wall or a bridge. A small inflatable boat may produce only a weak radar echo, which may be obscured by waves creating sea clutter. Rain creates unwanted returns, and land and port structures produce persistent reflections. Effective processing is therefore essential.Radar processing software first acquires radar video data through a hardware interface or a digital network stream. It then applies filtering and detection techniques to find returns that stand out from the background.This is more involved than simply setting a single threshold above which a radar video return is considered to be a contact. The background changes with range, platform motion, weather, and sea state. Settings suitable for calm water may produce many false detections in heavy rain; settings that suppress them may hide a genuine target.Adaptive processing responds to local conditions. Dynamic thresholding adjusts to changing noise levels, while geographic masks reduce processing over known land or structures. Where software allows, different radar processing settings can be applied in different regions, allowing greater sensitivity in a channel or directly ahead of the vessel. The goal is useful detections without excessive false alarms.Detection, tracking, and obstacle reportingOnce a significant return is identified, adjacent significant returns can be grouped together into a plot (detection). A plot is a single observation: something of interest appears to be present at a particular position during one radar scan.Tracking adds history and trajectory. Tracking software compares plots between successive scans and decides which are likely to belong to the same object. From these linked observations, it estimates course, speed, and likely future position. Confidence tests help prevent unrelated detections from becoming a false track.For small targets, tracking may become difficult as detection strength and reliability reduce. This is where tracking from video can help; a track that has already been acquired by traditional detection and tracking methods can be maintained for longer by tracking directly from the radar video.For autonomous navigation, it is not always appropriate to wait for a fully established track. Short-range or intermittent detections of objects such as a floating container or a small mooring buoy may generate a hazard that needs to be reported before a track can be generated. Proximity detections from a radar processor report the closest returns meeting configured criteria within a defined area around the vessel.Proximity detections are reported with very low latency, on each scan, as the radar turns. Safety areas can reflect speed, maneuverability, and environment. A forward zone might extend farther at higher speed, while a tighter zone may suit harbor maneuvers.Cambridge Pixel has been providing technology components for USV and ASV developers for over a decade, including its SPx Server V2 software, which includes radar video processing, proximity detection and tracking for collision avoidance and autonomous navigation. Its Detection Server configuration provides plot extraction and proximity detection, while the Tracking Server adds multi-hypothesis target tracking. Obstacle and target data can be supplied to an external autonomy or vessel-control system using formats including ASTERIX, NMEA-0183, TAK, and SAPIENT.Combining sensorsA radar track may show that an object is approaching or on a steady bearing without identifying it. AIS may provide identity data that can be correlated with the track, while a camera may offer visual confirmation. Sensor fusion combines these observations into a more complete picture.Successful fusion depends on consistent timing and coordinate systems and an understanding of the accuracy of each source, particularly where position is concerned. Bounding boxes defining uncertainty margins are critical with the aim of maximizing the number of contact correlations while minimizing false matches. Robust fusion is what turns multiple sensors into a coherent navigational picture.A correlated track report links the individual sensor observations that support it, allowing the receiving system to choose information from any of the inputs. For example, speed and course can be taken from the radar track, object classification and enhanced bearing measurement from a video track, and cargo information from the AIS.Cambridge Pixel's SPx Fusion Server correlates track reports from multiple radar sources and AlS into single fused tracks. Credit: Cambridge PixelPosition integrity without GPSAutonomous vessels commonly depend on GNSS/GPS. Loss of signal is a problem, whereas jamming can prevent reliable reception and spoofing can cause a receiver to report a plausible but false position.In coastal areas, radar can provide an independent cross-check of GNSS-derived positions. Cambridge Pixel's GPS Assist compares live radar data with a predicted radar image generated from terrain and coastline information. By finding the alignment that best matches the two images, it estimates the vessel's position independently of incoming GNSS data.If the positions differ significantly, GPS Assist can raise an alarm and also generate an alternative NMEA-0183 stream to provide backup positional information. Safe and effective maritime autonomy will not come from one sensor or algorithm. It will come from systems that provide independent, timely, and understandable evidence. Radar is central to that picture, but raw echoes are not enough. Only through rigorous processing can radar deliver the detections, tracks and obstacle warnings that autonomy systems depend on.