Naval asymetric threats and Piracy acts: study of a new IR detection module MALESE Dominique SAGEM Defense and Security Massy, France dominique.maltese@sagem.com HAMROUNI Ahmed ELECOM SUDPARIS Evry, France ahmed.hamrouni@telecom-sudparis.eu Abstract the latest events of naval piracy acts against commercial ships in the Gulf of Aden near Somalia have revealed a new threat for commercial ships. For military ships, naval asymmetric threats become also a more and more emerging threat to counter. Consequently, the implementation of protecting systems to face these acts (soft kill/hard kill actions according to civilian/military applications and contexts) turns out to be now inevitable. In a near future, commercial ships that are sailing in blue and brown waters in dangerous areas (straits ) will have to possess detection and countering capabilities to face single and multiple coordinated attacks from pirates in different short-range scenarios and environments (open sea, coastal backgrounds ). For military ships, the need is quite similar even if the threat and the dedicated reaction may be quite different (hard kill action if necessary). In this paper, a practical example of a module focused on the surveillance of the neighboring of the ship and its mean to detect and track such naval threats is discussed. he surveillance module presented hereafter is under study at SAGEM Company. he paper puts forward the detection process (image processing and tracking). he module may be combined with other ones to provide the ship with full range and angular coverage in its vicinity. o conclude, results of detection scenarios are provided. Data have been registered during sea trials dedicated to very short range threats. Results highlight how the module detects them. hey also put forward how module performances are improved by implementing two specific processes: track fusion and clutter map management. Keywords: tracking, image processing assymetric threat, piracy, uncooled InfraRed, cooled InfraRed, bandwidth, track fusion, clutter map. I. INRODUCION Nowadays, commercial craft have to face more and more piracy acts in specific well-known maritime ways such as Gulf of Aden. For instance, in 011, there were around 439 recorded incidents of piracy and armed robbery. Forty-five vessels were hijacked, 176 boarded and 113 fired upon. Last but not least, 8 crew members were killed. In the same way, naval asymmetric threats facing warships turn out to be a more and more growing target class to focus on. In open sea, attacks are expected to come from both the surface and below water by intruders attempting to attach an unidentified object to the ship. Intruders might also be fast boats with offensive weapons onboard (see Fig. 1). o illustrate these scenarios, in January 008, near the Strait of Hormuz, 5 fast boats, suspected to belong to the Iranian navy, raced by the wake of 3 US warships to harass them (USS Hopper, Port Royal and Ingraham). hey maneuvered in a hostile way at very short range from them (less than 00 meters away from the US Destroyer Hopper). he same scenario occurred in early January 01. Figure 1. example of naval assymetric threats (fast boats racing by the wake of a US warship) In order to be countered, these new naval threats have to be detected at short ranges (a few kilometers), and even at very short ranges (a few hundred of meters). Concerning piracy acts, since a few years, different solutions are applied to defeat such actions. he primary ones are: he presence of different navies in commercial areas where risks are high or increasing drastically. For instance, in Gulf of Aden, a task force of different foreign navies ( Combined ask Force 151 ) is being sailing to prevent pirates from hijacking ships (a real challenge when considering around 0,000 ships traveling these waters annually). he use of private companies to ensure the safety of the ships and their crews. Another solution under study is to set onboard ships, builtin sensors suites (mostly navigation radar, infrared/visible cameras distributed on the ship to cover all angles of attack, IR panoramic scanning equipments) that are coupled with non- 17
lethal technologies in order to detect, track, classify and defeat pirates with soft-kill actions (blind with light, drench with water cannons or deafen with sound blasts). he full set of data processing and acting steps have to be achieved in day/night conditions and different contexts (sea-state, blue / brown waters). hey are planned to provide a picket line to prevent pirates from approaching commercial ships. Concerning naval asymmetric threats associated to military ships, specific sensor suites are also elaborated to provide target indications at short range for soft-kill or hard-kill actions (use of fire-control systems). Generally, several sensor modules are installed at various locations on a ship in order to provide comprehensive 360 surveillance from the ship s side in the immediate and nearby vicinity. Panoramic scanning equipments are also used. Some major companies are now working on specific systems to face these new threats and make sea missions more secure. Of course, according to military or civilian applications, situations to deal with (and targets to counter) are not the same, what leads to different systems. Nonetheless, the use of quite similar modules is advantageous to benefit from companies experience at sea provided by current naval surveillance and fire control systems. hese modules can be combined between them to fulfill the system requirements (for instance, field of Regard of 360 in Azimuth). he use of Infrared technologies (cooled or uncooled according to requirements) turns out to be a good candidate. he resulting data can also be combined with other sensor data (radar ). In this paper, a practical example of a detection module studied at SAGEM is discussed. It puts forward the detection process (image processing and tracking based on Infra Red input data). Soft kill/hard kill actions (strategies on how to react according to data interpretation) are not focused. IR data may be combined with other sensor output such as radar data to: Upgrade output information quality and accuracies, Reduce system false alarms (alerts), Deal with possible lacks of detection from sensoralone platforms (induced by sea-clutter for instance). he detection module takes advantage of SAGEM experience with naval IRS comprising VMB and EOMS NG equipments ([3]). he last generation of EOMS NG is also under analysis to improve its capabilities when facing these new specific threats (see Fig. ). Section deals with the sense processing design. he sensor suite corresponds to uncooled IR cameras. Nonetheless, cooled IR cameras may also be implemented. Section 3 stresses on the sense processing functions that compose the system under study and more specifically the image processing and tracking functions. Processes provided in this paper for fixed sensors may also be applied as feedback for current panoramic scanning IRS equipments. Section 4 presents results on true recorded data. Adaptive aspects of the function and data qualities are put forward. Figure. IRS EOMS NG view & Zodiac boat tracked by EOMS NG II. SENSE PROCESSING DESIGN he Sense function presented in the paper is composed of 4 synchronized Uncooled IR cameras detecting in LWIR (8μm- 14μm see Fig. 3 and Fig. 4). Current Field of Regard (FOR) is 110 x40. So, it covers a wide solid angle in order to detect and track multiple targets (intruders) in a large angular sector. In order to cover the 360 azimuth coverage while handling masking effects onboard (blind arcs from masts, antennas ), 4 blocks have to be settled at different locations on the ship. Pixel size is around 1mrd so as to have enough detection capabilities for a surface target moving on a sea background. his last one may be at a few km or a few hundred of meters from the ship, according to the scenario to deal with (see Fig. 4). Figure 3. View of the built-in NIR Detection module (4 cameras inside) he current camera scan rate is more than 0Hz to reduce the lock-on time on the threat while having a very low system false alarms rate. Nonetheless, lower scan-rates (around 10Hz) and modified pixel sizes are available according to different applications or contexts of detection while keeping reasonable performances requirements. In the paper, uncooled IR technologies are selected (see Fig 4.). According to system requirements (accuracy, detection range ) cooled technologies may also be used in LWIR or in MWIR (3μm-5μm see Fig. ). he elevation FOR may also be modified. Figure 4. uncooled IR Image of an incoming zodiac at a hundred of meters from the sensor Detection module 18
III. SENSE PROCESSING FUNCIONS A. Introduction Sense processing functions are separated into three main functions (see Fig. 5): - Image processing : dedicated spatial filters are applied on the image to provide a first list of detection ( pre-alarm pixel ). At this step, no grouping action is done. So, various detections may correspond to the same extended target. For instance, this case occurs for threats close to the ship (range < km). Grouping action (labeling) is achieved later in the tracking module to balance time-running aspects between hardware and software implemented functions. he pixel data is segmented by managing a Horizon line that is estimated through inertial data. Image segmentation is done according to three main backgrounds: sky/horizon/sea. Other backgrounds and line of horizon profiles may be handled as well (short-range sea backgrounds, coastal backgrounds ). In these dedicated applications, sea background corresponds to the main background in the image or is the only background to deal with. - racking processing : first, pre-alarm pixels are labeled and grouped to provide a list of contact reports ( one report per target ). hen, this new list of data is handled by the tracking process to manage system tracks (creation/updating/deletion actions). - arget classification & output formatting : a threat evaluation is done by analyzing different extracted features associated to tracks (target size, range evaluation, ime- o-go (G), track behavior). argets of interest (bona fide tracks) are sent to the next module to process (visualization, combat system ) for confirmation and action to achieve. he module provides tracks with a certain level of confidence and that correspond to suspect tracks. Algorithms are based on evidential theories rules. On the following chapters, the paper will not focus on this specific function. Algorithms that are applied are mostly based on current and proven algorithms that are under use in SAGEM IRS equipments at sea (VMB, EOMS NG). he tuning of the parameters according to backgrounds and targets features has been achieved to fit a first roster of raw requirements. It makes the system able to detect and track naval asymmetric threats and pirate craft. he tuning will have to be updated according to other specific contexts of detection (coastal backgrounds) and threats (new speed boat capabilities, classes of pirate boats). Figure 5. main data flow of sense processing functions B. Image Processing According to the camera scan-rate, Image processing is integrated on FPGA modules (hardware) or on software components (CPU implementation). It is triggered at the frame rate of the synchronized cameras set. Its mains input are IR images & own ship attitude. Its output is the list of detection (pre-alarm pixels see Fig. 6) that is directly forwarded to the tracking process. Figure 6. main data flow of the image processing function wo main steps are applied: 1. hree spatial filters are implemented to deal with: Long-range detection (pin-point target detection), Short-range detection (targets entering the FOR close to the system), Intermediate range detection (targets entering the FOR at an intermediate range from the system). For each pixel & filter: Background statistics are estimated on the neighborhood of each pixel. Own ship attitude is not taken into account to achieve this task. Indeed, algorithms are enough robust to pitch, roll and yaw values evolutions. Pixel detection is declared if the pixel level exceeds a threshold. Detection thresholds depend on: he filter under use (main assumptions: pin-point target, large target and intermediate size target). At this step, different specific analysis functions are also applied separately for each filter to improve its own performances when facing cluttered or multimodal environments (clutter suppression, improved detection). hey are not provided in the paper. he position of the pixel according to the line of horizon (see Fig. 7). his data is estimated by taking into account the system attitude provided by the INS (pitch, yaw, roll and altitude). he transition area deals with uncertainties attached to the horizon line estimation (instrumental errors, sensor errors positioning). So, the horizon line is viewed with an angle that depends on the attitude of the ship (pitch and roll). 19
his way of doing reduces the image processing load. Nonetheless, Image Processing integrates this hypothesis (no image stabilization and preprocessing) by applying isotropic filters. It is reminded that in these specific applications under study, the major background (and in some cases, the only one) corresponds to the sea background. using INS data in order to track targets in a better way and for target kinematics interpretation. - then racking is processed (see Fig. 8): 1. First, confirmed tracks are addressed. hey correspond to tracks with a certain level of confidence ( track age threshold). then, hypothesis tracks are addressed ( track age < threshold) 3. remaining contact reports are at the origin of new hypothesis tracks (reports associated to none track) Figure 8. main architecture of the tracking process Figure 7. principle of Horizon line estimation. After spatial filtering, pre-alarm pixels that result on the three channels are merged and labeled. he Pixel Detection roster is then forwarded to the overall tracking process. Specific tuning has been applied to address threat detection ranges and false alarm rates under requirements. C. racking algorithms he tracking process is integrated in software modules (CPU implementation). It is triggered at the IR cameras frame rate. racking parameters depend on the background on which the target is detected (the background feature is obtained during the image processing by managing the horizon line). Its main inputs are the list of pixel detection sent by the low level Image processing. Output corresponds to the list of system tracks that is forwarded to the associated modules classification and output formatting. Main steps of the function are the following ones: - Pre-alarm pixels are labeled then merged according to a connexity principle in order to have the list of contact reports. More precisely, pixels on extended targets are grouped to have only one report per target (strong hypothesis for the tracking algorithms). During this step, associated features are estimated as well (average level, background information ). Resulting reports are converted into North East Down (NED) coordinates by During the management of confirmed and hypothesis tracks, specific algorithms are processed: 1. Reports and tracks are first clustered according to criteria of angular vicinities in order to be more efficient in time execution for next steps to process. Indeed, data assignment problem for large dimension matrices is broken down into separate data assignment problems for smaller dimension matrices (see Fig. 9). During this step, the predicted angular windows of tracks are handled to select reports that are close enough to them (predicted window areas correspond to areas where targets are supposed to be). hese gates are updated when the data association process is over (step 3). Figure 9. clustering principles. For each cluster of data, track-report assignment is solved by applying probabilistic data association techniques that manage multi-target multi-track assignments problem. For this application, NNJPDA (Nearest Neighbor Joint Probabilistic Data Association) has been used. Other assignment methods are also available (Munkres ). 0
3. racks are then updated at the current frame date: racks with any detection during a too long duration are suppressed in track pool (track deletion process). racks associated to reports are filtered according to a Kalman Filter method. he current model under use is a Singer model that supposes the acceleration has being a colorednoise process with a standard deviation σ m and an acceleration constant time τ m (τ m = 1/α). he model is well adapted for this specific context. Expressions of both transition and maneuver noise matrixes (F and Q) associated to the maneuver model are given hereafter: 1 F = 0 1, Q = ασ 0 0 1 m 5 0 4 8 3 6 4 8 3 3 6 Multi-model filters have also been implemented so as to be applied later for further applications. Angular windows are also predicted to the next frame date in order to make easier the data clustering to come. 4. racks are then fused according to behavioral similarities. More precisely, tracks are merged when they are close enough to each other (assumption made on a target size at a specified range) and when their kinematics states are quite similar. his way of doing is important to suppress multiple tracks on a same target track. Consequently, system performances are improved in numbering the targets present in the FOR. A clutter map has also been applied to deal with non homogeneous repartition of false alarms. More precisely, FOR is segmented into a grid of elementary angular areas expressed in NED coordinates (stabilized position reference). For each of them, dedicated features are assessed on the fly (average number of detections ). hey ensure to adapt initialization and deletion criteria according to resulting local detection densities. hus, clutter strength is handled here by delaying track confirmation process in accordance with system requirements (maximum time of initialization). his function is important to make the system self-adaptive with surrounding backgrounds variability (day/night conditions, presence or not of solar glint (sea-state )). After theses steps, tracks are forwarded to the module classification and output formatting. 3 IV. 1) Description of the scenario RESULS In this chapter, we assess the performances of the function whose input data correspond to true record data obtained during sea trials. he goal here is to quantify its performances on a head-on target in presence of sea clutter. Besides, ship wake phenomenon is also present in the scene. he target is not a pin-point target i.e. it is extended first on a small amount of pixels. hen, its size increases during the scenario while it is heading on the ship. Most of false alarms observed in the FOR are due to: Solar reflections on sea (low false alarm rate on sky) Ship wake generated by the threat. he tracker function under study manages detections that correspond to time stamped angular positions (azimuth & elevation). In this study, no morphological & radiometric features are taken into account in the association process. he interest here is to have a first overview on the system performances and more specifically on the tracker performances with only information even if these results may be upgraded later by adding new pertinent features in the correlation process (morphological/radiometric attributes). Be that as it may, the scenario that is studied in this paper gives a first trend on the tracker performances. hey have been confirmed by a more thorough study on a larger pool of scenarios. he intruder is an incoming Zodiac at a range of around 500 meters. he scenario duration is 3.4 seconds he maximum lock-on time on target is 1 second (tracker aspects) he camera data are: ) Results Pixel size ~ 1mrd camera scan-rate = 5Hz a) rack fusion Concerning the rue arget Detection & tracking aspects, the Zodiac is tracked from frame 10 to frame 810, i.e. during 3 seconds. Due to the track fusion module that is implemented, the function provides one track on the intruder though Image processing provides multiple detections on the target whose size is still increasing in the scenario (multiple contrasted points on the target). In fact, multiple tracks on the Zodiac have a quite similar trajectory what helps rack fusion module to merge them in a unique entity (group tracking). In order to illustrate the results, the Fig.11 puts forward more than 4 target tracks on the Zodiac when no rack Fusion process is activated. At the opposite, when track fusion process is triggered, Fig. 10 highlights only one track on the target of interest what is helpful for the operator. 1
Concerning the three tracks on the sky: - rack 1 corresponds to a malfunctioning pixel ( dead pixel) that provides a confirmed track. In the system, it may be strapped by a high level analysis of tracks during factory tests, or by the operator through a specific interactive visualization module (new malfunctioning pixels observation). - the two other tracks are obtained on a contrasted cloud edge Case Number ABLE I. SAISICS OF HE NUMBER OF RACKS PER FRAME Statistics of the number of tracks per frame Number of tracks per frame on Standard average the Zodiac deviation Case 1 no track fusion 1,95 1. Case track fusion applied 1,09 0.34 b) Clutter map he tracking process supposes the presence of a weak clutter. In the scenario, strong clutter is mostly present on the sky background (see Fig. 1). In order to reduce clutter effects, a clutter map has been implemented. It makes the tracker adaptive according to local detection densities (it is reminded that FOR is separated into elementary angular areas see previous chapter). Figure 10. example of results when track fusion process applied Figure 1. example of results when clutter map not applied able II. and III. put forward impact of clutter map in false tracks reduction. So, they are reduced by a factor 3 considering their average number. Besides, the associated standard deviations of their statistics decrease in a magnitude of. his gain improves system performances. Benefits of clutter map are quite similar when considering the full scene including sky and sea backgrounds (see able II) or only sea background (see able III). In this sequence, sea clutter (calm sea) is around ten times less important than sky clutter (presence of contrasted clouds). ABLE II. SAISICS OF HE NUMBER OF RACKS PER FRAME FOR SEA AND SKY BACKGROUNDS Figure 11. example of results with track fusion process not applied able I. provides statistics results on the number of tracks associated per frame to the Zodiac. It put forward a mitigation of the average value in a magnitude of with a good numbering of it. he standard deviation strongly decreases as well. In both cases, Zodiac target is tracked with the same duration. Case Number Statistics of the number of tracks per frame Number of tracks creation per Standard average frame deviation Case 1 no clutter map 0.16 0.4 Case Clutter map applied 0.05 0. Case Number ABLE III. SAISICS OF HE NUMBER OF RACKS PER FRAME FOR SEA BACKGROUND Statistics of the number of tracks per frame Number of tracks creation per Standard average frame deviation Case 1 no clutter map 0.017 0.13 Case Clutter map applied 0.007 0.086
o illustrate the results, only two false tracks are present in Fig. 13. heir number is 5 in Fig. 1. Figure 13. example of results when clutter map applied c) Conclusion Both of sub-modules (clutter map management and track fusion) are not yet optimal. hey can be highly improved. Nonetheless, they put forward a promising solution. hey largely increase robustness of the tracking process by reducing the clutter and avoiding one target to be followed by several tracks. Another result is that tracks following real target (not false targets), have a much longer time life. In other words, interruptions in target tracking (locks-off) are mitigated. he combination of them in the detection process is very important to provide system tracks to the situation assessment module. V. CONCLUSIONS Nowadays, commercial craft have to face more and more piracy acts. Different solutions are currently applied to solve this growing specific threat. One of them is to equip ships with dedicated sensor suites and non lethal protecting systems. Naval asymmetric threats become also a growing problem for military navies. In both cases, new dedicated equipments have to be achieved. hough they will surely be designed in different ways, the use of existing detection modules that benefit from proven experience at sea is interesting. Currently, concerning surveillance aspects (detection, tracking and classification of the targets), SAGEM is working on solutions based on IR cameras located at different spots on the ship to cover the full 360 azimuth FOR (collocated detection modules or not). Sensor data may depend on system requirements (accuracies, detection range ). Output data may be combined in a second step with other sensor output as those provided by the radar to improve the full picture of the scene. Algorithms that are used are based on SAGEM know-how in maritime detection environments (IRS, fire control systems). hey can be applied for both fixed sensor suite and panoramic scanning IRS. First results highlight good performance that may be still improved. New scenarios configuration are also to be added. Be that as it may, they put forward a promising solution to detect asymmetric threats and piracy acts. he function under study and more especially the tracking process, that includes the two extra modules, has also been tested on airborne scenarios. Resulting performances are even better than those obtained previously in naval context. his conclusion emphasizes how dependent of the environment the process is. It highlights the usefulness of both modules that are integrated in the tracking process. hey make able acceptable performances in different backgrounds and contexts. SAGEM benefits from experience at sea of its naval equipments and especially from IRS EOMS NG ([3]). Its current works on this detection module are on the improvement of its dedicated image and signal processes, especially on track fusion and clutter map management in order to: - keep avoiding multi-tracking on the same target while having a more steady ID on it, - reduce even more false tracks, - But also to detect targets earlier and further to give the craft more reaction time when facing these hostile actions. o conclude, SAGEM experience on naval IRS and its analysis on new threats to face, make it able to adapt its solutions to these new emerging and specific targets (use of fixed sensor suite or panoramic scanning equipments). [1] S. Blackman, R. Popoli, Design and Analysis of Modern racking Systems, Artech House Publishers (009). [] Y.Bar-Shalom, racking and data association, Academic Press Professional, Inc., San Diego, CA, USA (1987). [3] D.Maltese, P.O.Nougues and Al., New generation of naval IRS: example of EOMS NG, SPIE Vol 7660 May 010. [4] D.Maltese, A.Lucas, Data Fusion: Principles and Applications in Air Defense, SPIE Vol 3374 April 1998. [5] P-O. Nougues et al, hird-generation naval IRS using the step-andstare architecture, SPIE Defense & Security 008. [6].Kirubarajan and Y.Bar-Shalom, Probabilistic data association techniques for target tracking in clutter, IEEE, 004, pp 536 556. 3