IntroductionRelated WorkSEI SpecificComputer VisionOur ModelAssumptionsNoiseMaritime WeatherNuisance ArtifactsGeneralized ApproachVideo Tracking and StabilizationKLT Feature ExtractionTranslation Estimation with LLSTranslation Estimation with CentroidTranslation Estimation with RANSACPeriodic Signal AnalysisCutler and Davis MethodSimpler MethodScan Line PositionConclusionReferencesAbstract The ability of an unmanned system to visually determine the direction a ship’s radar antenna is facing can significantly improve the SEI collection process and provide opportunities for autonomous collection. This paper investigates methods for detecting and analyzing periodic motion of spinning radars in the maritime domain. The methods include feature based tracking and periodic measurements using similarity matrices. 1. Introduction SEI collection is an important asset for information dominance. “SEI provides a reliable, long-range, all-weather positive target identification capability against seaborne platforms and land-based systems that emit radar signals.”[1] The process of SEI collection is to “find non-intentional modulations in the receiving signals”[2]. An illustration of the required components is shown in Figure 1. Figure 1 SEI Collection Cartoon [4] The SEI collection procedure currently requires a human-in-the-loop. The operator (human-in-the-loop) is required to point the SEI collection antenna at a ship’s spinning radar using optics bore-sighted with the collection antenna. The operator (sometimes a second operator) configures the SEI collection tuning parameters while the optic system is manually positioned to follow the ship radar. An SEI collection occurs when the operator confirms the position (shown in Figure 2) of the radar antenna is directing its beam at the collection antenna when the SEI system signifies a signal has been received. The collection is complete after the operator reads the name of the ship, confirms the radar signal is indeed from the ship, and acquires other distinguishing features (color, hull type, etc.) Figure 2 The angular position of radar we want to detect The procedure the operator employs is an important step for providing ground truth. One of the four key core concepts for an SEI system design is “Providing ground truth (the correct identification of the emitters being evaluated) for the naming of the clusters and evaluation of the clustering process.” [2]. This is a complex, error-prone, and time-consuming procedure which could be significantly optimized through automation. The automation is addressed by researching methods for radar angular position and periodicity verification. For most collection sites, this step alone will cut down the need of two operators down to one. The efforts addressing this vision system concept can also be extended to provide cueing for ship detection, tracking, and identification. Another reason for autonomy is collection in hostile environments. SEI collection using the current procedures is not a safe option because of the risk of placing human operators in hostile environments. This risk limits the ability to monitor vessels of interest (VOI). A VOI can avoid intelligence collections by staying within the hostile environments. An autonomous SEI collection system would best support this scenario and could be tremendously advantageous to the DoD mission of MDA. Radar Antenna Detection and Analysis for Autonomous SEI Collection Kris Gibson Department of Electrical Computer Engineering University of California, San Diego [email protected]. Related Work 1.1.1 SEI Specific There are a multitude of research journals that address the electromagnetic signal parameter realm of SEI. These methods employ pattern recognition, classification, and feature extraction of the received signals [2]. 1.1.2 Computer Vision At this point in time, there is no other research using a computer vision approach to detecting the period and angular position of a radar on a ship. Work that is closely related to this topic includes tracking and stabilization methods, and periodic motion estimation. There exist many different video stabilization systems that are hardware and also software solutions. Some use affine transformation models as in the Motion2D software library from Inria. There are feature based methods which we will explore in this paper provided by a Kanade-Lucas-Tomasi method. There is a vast amount of literature focused on the aspect of detecting or recognizing the periodic nature of rigid and non-rigid moving objects. In general the objects of interest are biological such as humans, animals, and trees [5]. We will follow closely the work presented by Cutler and Davis [7] which uses a similarity matrix representation to measure periodicity. 2. Our Model 2.1. Assumptions The primary object of interest is rotating radar mounted to a ship. Our model assumes the viewing distance is long enough that an affine transformation model can satisfy all dominating transformations. This has been verified by measuring corresponding points of a ship in sample video sequences. In fact, the model has shown a dominant translation factor and very minimal rotation, scale, and skew factors. It is also assumed the field-of-view (FOV) is narrow enough to be able to view the radar. If the rotation of radar is not distinguishable from other motion artifacts in the video sequence then this model will fail. It is also assumed for this effort that the radar is mostly in the FOV during the whole video sequence. The length of time the radar must be in the FOV is dependent on the slowest periodic rate of rotating radar. A typical range of rotation speed is between 24rpm to 50rpm. This means the slowest period would require a 2.5 second window of time the radar must be in FOV. The radar may be partially occluded. The number of radars that exist in the FOV is not limited. Typically multiple radars on a ship will have a different periodic rate and position. This model relaxes the requirement of number of ships in the FOV. There are many reasonable