ASTRA: ACTIVE SHOOTER TACTICAL RESPONSE ASSISTANT ECE-492/3 Senior Design Project Spring 2017

Transcription

1 ASTRA: ACTIVE SHOOTER TACTICAL RESPONSE ASSISTANT ECE-492/3 Senior Design Project Spring 2017 Electrical and Computer Engineering Department Volgenau School of Engineering George Mason University Fairfax, VA Team members: Faculty Supervisor: Ben McCall, Puja Patel, Rohini Shah, Aryan Toughiry, and Joel Williams Dr. Kathleen Wage and Dr. Kenneth Hintz Abstract: ASTRA is a network of acoustic sensor arrays that provides law enforcement, private security and emergency medical teams with fast and accurate location information in an outdoor active shooter situation within seconds of a shot being fired. ASTRA is an automatic detection system that is not dependent on any human input, which can be prone to error in a chaotic, high-pressure situation. ASTRA is able to detect a list of specific firearms, which was determined after researching which specific firearms were used in active shooter situations. 1. Introduction According to the United States Federal Bureau of Investigation (FBI) there has been an average of 11.4 active shooter incidents per year since 2003 in the United States, and the trend has been increasing. In the last seven years the average number of incidents has risen to an average of 17 incidents a year. We can conclude that active shooter situations are a major concern in our country. According to the FBI there have been a minimum of 1,043 casualties since From there have been 160 active shooter incidents and almost every incident has involved a single shooter. Since 2003, 45 people have died and 54 people have been wounded in active shooter incidents that were located in an outdoor open space, which is specifically the problem statement that ASTRA will solve. Developing ASTRA would be beneficial by first, getting quicker and immediate responses to the civilians who were wounded and second, by locating the shooter more quickly and precisely than previously used methods. Two major competitors to the ASTRA system are Shotspotter and Boomerang. Shotspotter is a nodal network gunshot detection and localization system that relies on cabled sensors that connect to a main operations center somewhere in the geographic locale where it is deployed. This system relies heavily on infrastructure and because of its reliance on time synchronization between the nodes as a function of path delay, it is therefore not mobile. ASTRA has GPS on each node, which is updated at the data receiver, which computes a geographic approximation of the shooter s location based on independent node reports without having to synchronize the timing of each node. The Raytheon Boomerang system is a single node system, which provides military personnel with direction and distance of a shot, however the round must pass within 30 meters of the node. ASTRA uses a multi-node approach to achieve a precise distance to detection and each node can plot the angle of a shot at a distance of over 100 meters. Because ASTRA can be deployed using large amounts of nodes, it has the potential to provide greater coverage and higher accuracy than Boomerang 1

2 2. Approach ASTRA is a system that uses direction-finding nodes. Each node independently directs a vector to the shot. Match filtering is used to construct a set of predetermined filters that detect a gunshot. The acoustic signature of a gunshot is recorded and then convolved with each of the matched filters in the system. The convolution output shows a peak at the location of the gunshot and if the threshold criteria is met, the system will report a positive detection and advance to the next stage, localization. After the initial detections, each node uses cross-correlation to perform direction finding which determines a specific direction from which the gunshot originated. Each nodes angle is different because they are calculated with respect to geographic position relative to the gunshot. Each node wirelessly sends a TCP packet containing 4 essential elements (angle of arrival, time of detection, node number, GPS coordinates) back to the data receiver. The data receiver then takes each nodes information and displays a region in the ASTRA terminal that shows the approximate location of the shooter as shown in Figure 1. Figure 1: Detection, direction, and location finding of a shooter 3. System architecture The design of the ASTRA system consists of a network of directional finding nodes (three nodes in our case) and one data receiver. Figure 2 shows a high level diagram of ASTRA s general design. Node 1 Node 2 Listen Detect gunshot Timestamp detec on Report detec on to server Listen Detect gunshot Timestamp detec on Report detec on to server Node 3 Listen Detect gunshot Timestamp detec on Report detec on to server ASTRA Data Receiver 5/32 Figure 2: Overall system design (three direction finding nodes and one data receiver) Each node (Figure 3) uses 4 omni-directional condenser microphones connected to a 4-channel A/D converter that has 4 microphone preamplifiers and provides 48V phantom power to allow the condenser microphones to operate. After pre-amplification, the analog audio input undergoes A/D conversion, after-which the digital output goes through a detection filter. In the detection filter stage the digital signal runs through a bank of match filters based 2

3 on specific gunshot signatures. If the convolution output meets the pre-determined threshold it sends a positive detection flag to the next stage. If there is a positive detection flag the phased array calculation stage applies the concept of time distance of arrival to calculate the angle of the gunshot relative to the position of the array and the resulting angle is calculated. In the GPS/TCP packetization stage, a GPS antenna on each node is used to collect each node s GPS position, timestamp of the gunshot, angle to the shot, and node number. These 4 pieces of information are collected into a TCP packet and sent to the data receiver over a wireless backhaul. Figure 3: Single node architecture The final stage of the functional node structure is the wireless backhaul (Figure 4). Our system uses ISM 900MHz WiFi to act as the system s own independent network. Each node follows the steps described above and sends a TCP data packet to the data receiver. Our system consists of three nodes, which uses time distance of arrival processing methods to send a data packet via a TCP message to the data receiver. Afterwards the data receiver uses a JAVAprogram to generate a KML file, which contains three directional vectors, their uncertainty margins, and the circular error of probability, which represents the area where a shooter fired a shot. Figure 4: Flowchart of the data receiver's process 3

4 GNURadio (an open source digital signal processing software) connects to audio or RF hardware for real time processing of streaming data. GNURadio allows the user to utilize a wide range of traditional and user-created DSP operations in code blocks. These blocks when connected together create 'flowgraphs' that perform multiple operations. Figure 5 is an image of ASTRA's GNURadio software module. Figure 5: ASTRA's GNURadio processing module diagram 4. Theoretical background Matched filtering is used to match to a signal (given at s(t)), at some time T). The filter which is matched to the input signal will always have the larger output [1]. The impulse response of the matched filter is given by (1). h(t) = s (T t) (1) To create the matched filter, the gunshot is isolated within a specified time window, and it is then time reversed and shifted. FFT is used to improve efficiency of isolating the gunshot N 1 x[n] = 1 X[k] N k=0 e j(2π N )kn (2) X[k] = N 1 n=0 (3) x[n]e j(2π N )kn with a number of operations proportional to Nlog 2N rather than N 2, where N equals the number of samples [2]. For a 10,000-point transform, the FFT gives approximately 752-fold speed improvement. Instead of finding the 4

5 output using a traditional time domain convolution between the raw gunshot and the pre-designated match filter, a more processor efficient frequency domain FFT multiply is used. Two fundamental aspects to be addressed in automatic target detection and recognition applications are its ability to distinguish targets from background noise and its ability to recognize an actual target. CA-CFAR presents adaptive digital signal processing algorithms which adapt the threshold automatically to local information on the background noise. This is incredibly important for a signal processing system, to be able to operate in a nonstationary background noise environment while maintaining a constant false alarm rate, or a constant level of performance. A false alarm is when the system reports a positive recognition when there was either no recognition to be made or the system recognized an incorrect signal. By processing a window of reference cells surrounding a central cell under test, the CFAR algorithms are able to estimate characteristics of surrounding noise [3]. Through averaging the reference cells, the CA-CFAR detector regulates the false alarm probability to a desired level. From this, a decision for target present and target absent can be made through (4) using an adaptive method that takes environmental noise components into account in its identification of the sensor signal. Figure 6 shows the implementation of CA-CFAR on the output of a matched filter from the ASTRA system. e(y) = { 1, if Y T Y AC 0, if Y < T Y AC (4) Figure 6: ASTRA's GNURadion processing module diagram Once the gunshot is detected, the three spatial sensing microphones as well as the central detection microphone displays their delayed inputs. Auto-correlation is then used to find the time zero from which the other delays are found. The peak of auto-correlation is at a lag of zero. Then, the cross-correlation function is used to find the time delay [4] between the central detection mic and the three spatial microphones. The peak of the crosscorrelation is where the two signals are most aligned. The lag (measured in samples) is then found from the time zero point to the peak of each cross-correlation. To determine the direction in which the sound originated, also known as sound source localization, plane wave geometry was used. Assuming that the sound source is considered a long distance from the array, a plane wave 5

6 can be assumed [5]. This far-field approximation classifies how a military impulse event will propagate across the array as a uniform wave front [4]. To find the inaccuracy in the directional vector originating from each node, the vector is compared against another vector that originates at the same node, but instead points to the known shooter location. The difference between the two angles is denoted as alpha (measured in degrees) and the standard deviation of these values is used as a visual aid (denoted by dotted lines surrounding the directional vector) to show statistical analysis of the systems performance. Note that all angles are respective the coordinate system shown 5. Experimental Evaluation and Validation ASTRA testing process was a 6 step procedure in the following order: Benchtop Testing single node, Loudspeaker Testing single node, Live-Fire Testing single node, Benchtop Testing multi node, Loudspeaker Testing multi node, Live-Fire Testing multi node. We took this meticulous process of testing to make sure we did most our debugging in the lab (Benchtop testing) verses the field where we would have less resources. Once the software functions were working the preliminary bench-top test consisted of each node being connected and run as if it were in the field. The first purpose of the bench-top test was to confirm that each of the nodes accurately detected the gunshot and reported an angle. The second purpose was to ensure that the network reporting system worked correctly, so that we could see each node detect a gunshot, plot an accurate angle based on simulated delays, and report that detection and angle back to the data receiver. The third purpose was to test the data receiver s ability to handle each report from the ASTRA nodes and plot the node locations and angles of detection on the ASTRA terminal. Once our results for bench-top testing were validated by the ASTRA terminal window on a laptop showing three nodes with three directional arrays creating an intersection, we moved to loudspeaker testing which was simulating the gunshot through a loud speaker. Once our loudspeaker testing was successful, we moved to the final stage of testing, which was live fire testing (testing with actual gunshots at Naval Research Laboratory's Chesapeake Bay Detachment). Figure 7 shows a detection of a gunshot displayed on the ASTRA terminal. Figure 7: Post-shot display with confirmed GPS location for the location of the actual shooter Once we had confirmed our Live-Fire Test (our final test) was a success, we moved on to the analysis of our results which are shown in Figure 8. 6

7 Figure 8: System error results Finally, we took into effect all of system error results and applied the theoretical concept of Circular Error of Probability (CEP) to localize a confidence region in which the shooter shot their firearm, the radius of aforementioned circle was 2.14 meters. Figure 9 shows the final display the user would see when a gunshot is detected and localized. Figure 9: CAP on ASTRA terminal References [1] Ephraim, Yariv. ECE 460 Lecture Notes: Communication and Information Theory. 2016th ed. N.p., [2] McClellan, James, Ronald Schafer, and Mark Yoder. DSP First: A Multimedia Approach. Upper Saddle River, NJ: Prentice Hall, [3] Cumplido, R., C. Torres, and S. Lopez. On the Implementation of an Efficient FPGA- Based CFAR Processor for Target Detection. (ICEEE). 1st International Conference on Electrical and Electronics Engineering, N.p., [4] Rhudy, Matthew et al. Microphone Array Analysis Methods Using Cross-Correlations. (2009): [5] Greensted, Andrew. Delay Calculation. The Lab Book Pages. N.p., 2 Sept