Perimeter Security Intruder Tracking and Classification Using an Array of Low Cost Ultra- Wideband (UWB) Radars Henry Mahler, Brian Flynn Time Domain Corp Huntsville, AL Henry.mahler@timedomain.com Abstract Time Domain has developed a perimeter security system based on an array of its commercial off-the-shelf (COTS) UWB radios, which operate as radars in a mono-static or bi-static mode. The system consists of UWB radios within poles, a wired network delivering data to a server and software operating on the server. The wired network and server are constructed from COTS components. Much of the system is implemented in software making the system open and easily upgradable. As a radar based system, it can track and classify intruders in all weather. I. INTRODUCTION The perimeter security system provides detection and tracking of intruders attempting to pass through a virtual fence. Originally developed in conjunction with the Navy as the Shore Line Monitoring System (SLiMS) it can be deployed at any location requiring perimeter security. It detects, tracks, and classifies target(s) as they approach and penetrate a perimeter. Since the system can track targets beyond the actual perimeter boundary it provides awareness of activity in the area even if no attempted intrusion has occurred. Target classification is accomplished using an imaging algorithm to determine whether the target is a, animal, or. Track direction in conjunction with classification allows responders to quickly determine if a threat exists and the nature of the threat. With this information security personnel can respond in the appropriate manner. The system has been deployed at an operational site for over 2 years and has demonstrated a high probability of intruder detection, a low probability of false or nuisance alarms, and a high probability of correct classification. The system consists of a dual line of poles spaced approximately 2 meters apart with three Time Domain UWB radios, operating as radars, in each pole. The main components of the system include the poles, the wired network, and the server resident software. In addition to housing the UWB radios the poles also contain COTS networking gear that connects the poles to the wired network. Raw data collected by each radar is sent to a central processing server over Ethernet. The processing engine on the server detects, tracks, and classifies targets. A display application is used to show track locations and classifications relative to the area of interest as well as notifications of perimeter breaches and system status. The COTS UWB radio technology allows the radars to operate in both mono-static and bi-static modes. The very high range resolution of the UWB radars enables use of omni-directional antennas even in moderate clutter environments. A time division multiple access (TDMA) network coordinates which radars transmit at what times. The ability of multiple radars, operating in bistatic mode, to receive the signal transmitted from a single radar is key to exploiting the available target energy. The processing engine at the central server employs state of the art radar processing to detect, track, and classify targets. A motion filter is used to detect moving targets in clutter. A constant false alarm rate (CFAR) algorithm is used to reduce false clutter detections. A multiple target tracker is implemented using a Probabilistic Data Association Filter (PDAF) and Interactive Multiple Model (IMM) Kalman filter. Target classification uses a back projection imaging technique and a naïve Bayes classifier with a cumulative likelihood measured over multiple target updates. II. APPROACH A. Ultrawideband (UWB) Radar UWB radar transmits high bandwidth (narrow) Gaussian pulses. With a center frequency of 4 GHz and a bandwidth of 2 GHz (see Fig. 1) the waveform supports a resolution of a few centimeters. Because the transmitter is operating under FCC part 15 the power in the pulse is less than milliwatt [1, 2]. The UWB radios receive the pulse response in either a mono-
static mode or bi-static mode. Each radio s precision on-board timing allows it to operate as a mono-static device where it first transmits a pulse and subsequently receives the pulse response. In the bi-static mode, one radio transmits the pulse and another radio receives the response. In order to enable bi- an acquire static operation the transmitted waveform includes sequence that allows the receiving radio to find the transmitted pulses in time. Once found, a code embedded in the polarity of the transmitted pulses allows the receiving radios to synchronize with the transmitter. When multiple receivers receive the same transmitted pulse response this is termed the multi-static mode and as we will see this mode is fundamental to the system. Within the system a complete pulse response is known as a scan. The sequence of scans between a specific transmitter and receiver pair is a link. the average number of scans per second captured by the cell as a whole is about 432. The system also supports a feature where radios receive transmissions from adjacent cells. This occurs when a receiver in one cell is closer to a transmitting radio in an adjacent cell than it is to a transmitting pole in its own cell. This is possible because all cells sync to the same TDMA network. When a radio captures a scan, it is packaged into UDP packets and transmitted to a server via the wired network. 1.5 db -1 3.1 5.3-2 -.5-1 -1-5 5 1 t, pico seconds -3-4 1 2 3 4 5 6 7 8 Frequency, GHz Figure 1. Time and frequency measurements of the fundamental pulsed signaling strategy of the P4 radio module B. Radio Network Three UWB modules are mounted within each pole (see Fig. 1). The poles are positioned in two parallel rows spaced 2 meters apart; poles within each row are located 4 meters apart. Although all radios have a transmit capability currently only the middle radio on each pole transmits. With the limited transmit power available one transmitting radio will not be received by all radios in a system. To overcome this the system is divided into six pole groups which are called cells. The transmitting radios in a cell coordinate their transmissions by forming a TDMA network. Each transmitting radio has a slot in the TDMA network and all other radios in that cell listen during that slot. A six pole cell contains 18 radios thus for any given slot one radio transmits. It also operates in the mono- as bi-static static mode, and the remaining 17 radios operate receivers. Fig. 2 illustrates the case where the middle radio on pole 1 is transmitting. The number of links in a cell is 18. The cell can have more than 21 geometrically unique links. The three radios on each pole provide multiple views at a target. The object of interest (s, and s) tend to provide multiple complex returns to the radars. Sometimes, the multiple scatters may combine in such a way to cancel target responses to one of the radios on a pole but other radios on the pole will see a different response wheree the target is not canceled out. Another factor is ground bounce, the path between one radio and a target may be experiencing a null but the path to the other two radios will lie outside that null. The TDMA network cycle rate is approximately 4 Hz, thus each pole will transmit 4 times a second. With six poles in cell, Figure 2. Network- One Slott One Transmission C. Processing Engine One thing that differentiates the system from other systems is that the sensors, the radios in this case, perform no processing on the data. Instead the raw data are delivered to the engine running on the server. As shown in Fig. 3, the steps involved in processing that data are signal processing, detection, localization, tracking, and classification front end processing. Figure 3. System enginee block diagram a) Signal Processining Raw scans need a number of signal processing steps before the scans can be evaluated for detections. First the scans are checked for validity. Second, a band pass filter is applied in order to improve SNR. The radios do not always start capturing the scan at the exact same point. To compensate for this, later scans are shifted to align with earlier scans in order to minimize correlation residuals. Finally the scans are motion filtered against the previous scans from the same link.
b) Detection Detections are points in a motion filtered scan that are sufficiently different from data contained in previous scans from the same link. The detection processing uses a standard CFAR technique to find these changes. c) Localization At this point the system has a list of detections from some number of links. The next step is to evaluate this list of detections for information about targets in the area. There are many approaches that would work [3, 4]. The approach used is to convert from detection/link space to 2D coordinate space through a localization process and apply the output to a linear Kalman tracker. Due to the high resolution and high number of links a single target can generate lots of detections. The compute-intensive localization process distills a large number of detections to few cluster points. Detections on a link define an ellipse in 2D where a target could be. Note a circle can be parameterized as an ellipse, so detections from both mono-static and bi-static links are defined as ellipses. The localization process computes the ellipse intersections of the detections. This process is not quite O(N 2 ) because there is no need to compute crossing for detections on the same link, and the system skips computing intersections where the radios are too far apart to detect the same target. A pair of detections can generate up to four intersections. Each intersection has a signal strength metric computed from the detection of that intersection. Next, the localization process searches the intersection map for cluster points. The clustering process looks for the highest concentration of signal strength or energy. The center of the energy is chosen to be the cluster point. Then all detections associated with that cluster point are removed from the intersection map and the process is repeated. When either no detections remain or a fixed number of cluster points are reached the process is terminated. The process of computing intersections and searching for cluster points requires sufficient computational resources. The system maintains real time operation by using a computer with multiple cores and by limiting the number of detections processed. The localization process evaluates detections at the TDMA network cycle rate which is approximately 4 Hz. Fig. 4 shows the ellipse crossing and cluster points generated by a single person during one TDMA cycle. d) Tracking The next step is to apply the cluster points to a Kalman filter based target tracker. Each cluster point is provided as a measurement to the tracker which evaluates the cluster points both temporally and statistically to find sequences of cluster points that indicate the presence of a target. Figure 4. Localization from a single TDMA cycle The actual tracker uses a PDAF to associate measurements (cluster points) to existing tracks. Measurements that do not associate to tracks are used to spawn new tracks. The tracker uses a PDAF instead of JDPAF because measurements of two targets do not tend to interact to a significant degree. The main role of the tracker is to estimate states of the tracks. A four state vector [x, y, x, y], represents 2D position and velocity. To get improved estimates of the states, the tracker includes an IMM with two models, one model for faster moving targets and one for slower targets. The target generates a metric indicating the quality of the estimates. The metric has a higher value when a continuous sequence of measures is associated with a track and the metric declines when a track stops getting measurements. The tracks whose metric exceeds a threshold are forwarded to the UI for additional qualification before being displayed. e) Classification Front End Once a track meets certain qualifications, the system starts a classification process for that track. This is a sequential procedure where features of the tracked object are evaluated over time (over many track updates) until a determination can be made as to whether the object is, animal or. A naïve Bayes classifier is used to evaluate the feature. The first step is to generate a 2D image of the radar data using a back projection technique, see Fig. 5. The 4 meter square image (with a pixel size of.6 meters square) is centered on the track position. An example image of a is shown in Fig. 6 and the image of our simulator in Fig. 7. An alignment vector (a vector thru the longest part of the image) is derived from the track states. Four features from the image are evaluated, length of object with respect to the alignment vector, width of object with respect to alignment vector, angle between alignment vector and velocity vector and a shape feature (χ 2 ) indicating how elliptical the image is. Two other features derivate from the base radar data, one feature is peak amplitude and the other is target length in tau space. These features are used to differentiate between the three classes:, animal, and. For example, target length is valuable because animals tend to have a longer length than s. The angle between alignment vector and velocity vector is a critical feature because when the animals
p m ( ) = density function for classs m, m = 1,2,3 (3) The functions are computed numerically from tables. The tables are generated by evaluating known truth data contained in an extensive radar data database maintained by Time Domain. Fig. 8 shows the density functions in graph form. Figure 5. Imaging Back Projection p(aalgn) p(x2) p(amax) 2 1 1 2 3 4 5 6 Dlen (m) 4 2 1 2 3 4 5 6 Dwid (m).4.2 1 2 3 4 5 6 7 8 9 Aalgn (deg).4.2 1 2 3 4 5 6 X2.2.1 45 5 55 6 65 7 75 8 85 9 Amax (db) 2 1.5 1 1.5 2 2.5 3 3.5 4 Lseg (m) p(dlen) p(dwid) p(lseg) Figure 6. Image of Human Figure 7. Image of Deer Simulator are moving their velocity vector is within a few degrees of their alignment vector. Human walking tends to exhibit a velocity vector perpendicular to the alignment vector. Next, for each class (, animal and ) the system computes the log likelihood for this update. L m = -Σ ln(p m (f k )) (1) f k = feature k, k = 1,2,3,4,5,6 (2) Figure 8. Plot of Density functions, from top to bottom: length of object with respect to the alignment vector, width of object with respect to alignment vector, angle between alignment vector and velocity vector and a shape feature (χ 2 ), peak amplitude, target length in tau space The system then accumulates the values for each of the three possible classes, animal, over many updates. A decision on target classs is made when the system has evaluated enough track updates. The system does not stop evaluating track points when a decision had been made. The evaluation continues and if later data indicates a different class is more likely, the system can change its decision. The classification function is divided between two applications, the engine and UI/ODP. The engine does the computationally intensive portion of the processing because the engine is hosted on a server and the engine has access to all radar and track data. The second part of the classification process of actually accumulating the class values and making the decision of which class to declare is handled by the UI application. D. User Interface (UI) The UI is a separate application to display the results to an operator. The application shows active tracks color coded according to classification results. In addition the UI provides radio status to the operator; for example, an icon is displayed if a radio stops functioning. The UI maintains a log of system operation. This log can be used in conjunction with a replay feature to display historical events... Fig. 9 shows the output of the UI.
objects the Probability of correct classification has also been collected. The system was evaluated on a continuous basis throughout its development; the most comprehensive data set was collected from exercises collected at an actual deployment site in February 214. For Pd, the system detected a target on the each of 93 test tracks presented to the system. This was a mix of targets, s walking, running and crawling, s and simulator. The false alarm rate was on average less than one per day. Figure 9. System User interface III. CONSTRUCTION AND INSTALLATION The main components of the system include the poles, the wired network, and the server. Each pole is approximately 3 meters tall,.15 meters in diameter, consumes 3 Watts of power and weighs approximately 13.6 Kg. The three UWB radios are deployed at different heights within the pole. The Ethernet switch and AC/DC power converter are housed at the base of the pole. Fig. 1 provides an exterior view of a typical pole. Because Time Domain envisions a system as part of a site s permanent infrastructure pre-installation coordination with the site integrator or end user is required. Following installation the cost to maintain a system have proven to be minimal. The classification results are summarized in Table 1. The TCO means total classification opportunities, TCC means total number of correct classifications, and PCC is percent of correct classification. Each track was assessed at three points for classification. Each assessment is a classification opportunity. The 93 tracks produced 279 classification opportunities. The classification results from two s walking with less than 1 meter separation occurred because the system merged the two s into a single 2D image and this image did not display the characteristics of one of the current target classes (, animal, or ). In the future, the system will include a classification type of multiple s. In addition after examining the results for running s a change was implemented. That change is undergoing testing at this time. One aspect of the system not considered by normal performance metrics is the value to the operator of being presented with tracks of targets instead of zone alarms. There is not a defined metric for this. However operationally one can appreciate the difference in organizing a response to a zone alarm and organizing a response to a known target type at a known location. V. CONCLUSION The system offers unique features for a perimeter security system. The system relies an array of low cost sensors working together to monitor a section of a perimeter. That makes it less susceptible the foibles of a single sensor. The system provides an alternative to traditional trip line sensors that only provide zone alarms. Most significantly it provides track and classification information about a target(s) giving the security operator better situational awareness. Figure 1. Pole A set of system demonstration videos that illustrate the system s response to an array of scenarios is available at http://goo.gl/z63cw4. IV. PERFORMANCE The goal of any perimeter security system is to detect real intruders and not generate false alarms. The metrics associated with these goals are Probability of detection (Pd) and Probability of false alarm (Pfa). Since this system can classify
Table 1: February 214 Classification data, * See discussion under Section IV Target Tco Tcc Pcc, % Single walking 18 18 1 Two walkers ~ 1 meter apart* 24 Two walkers >= 3 meters apart 36 36 1 Single low crawling 15 11 73.3 Single bear crawling 9 8 88.9 Single running 12 5 41.7 Deer simulator 45 37 82.3 Vehicle 3 28 93.3 REFERENCES [1] A. Petorff, A Practical, High Performance Ultra-Wideband Radar Platform IEEE-AESS RadarCon 212 (Atlanta, Georgia May 7-11, 212) [2] Time Domain Web Site. Time Domain s Ultra-Wideband (UWB) Definiations and Advantages, June 212 [3] Yaakov Bar-Shalom and Xiao-Rong Li, Multitarget-Multisensor Tracking: Principals and Techniques YBS publishing, 1995. [4] Yaakov Bar-Shalom, X. Rong Li, TThiagalingam Kirubaran, Estimation with Applications to Tracking and Navigation: Theory Algorithms and Software 1985.