Bark Indication Detection and Release Algorithm for the Automatic Delivery of Packages by Dogs

Transcription

1 Bark Indication Detection and Release Algorithm for the Automatic Delivery of Packages by Dogs Jimmy Tran Ryerson University 350 Victoria Street Toronto, Canada M5B 2K ext Alexander Ferworn Ryerson University 350 Victoria Street Toronto, Canada M5B 2K ext ABSTRACT The Canine Remote Deployment System (CRDS) is a set of equipment for the delivery of emergency supplies to trapped victims in urban search and rescue (US&R) operations. The system utilizes search canines carrying a bag of supplies and equipped with a wireless receiver device. A hand-held wireless transmitter is used by a human to trigger the release of the bag. The canine handler would activate the transmitter upon hearing the dog s bark indication that it has found a live victim. In such manner, a package can be deployed from the dog at the press of a button. This paper describes a new feature of the CRDS whereby the release mechanism that deploys the package called the underdog is activated through the detection of the dog s barks and an algorithmically controlled deployment strategy. A series of experiments were conducted to ensure consistency and accuracy of the bark release algorithm. Results are presented in this paper. The Canine Remote Deployment System (CRDS) was developed to increase victim survival rates during the often long interval between discovery and rescue. As described in [3], CRDS utilizes the agility and determination of US&R dogs to deliver food, water, medical supplies and communications devices to victims trapped in areas where passage by human workers is difficult or not practically possible. The supplies are stored in a bright orange bag that the dog carries under its belly called the underdog as shown in Figure 1. Categories and Subject Descriptors B.7.1 [Integrated Circuits]: Types and Design Styles Algorithms implemented in hardware General Terms Algorithms, Design, Experimentation, Theory Keywords US&R; canine; automation; CRDS; Biological Intelligence 1. INTRODUCTION In urban disasters involving the structural collapse of large buildings, people may be trapped in voids underneath rubble [1, 2]. As a prerequisite to rescue, US&R first responders must locate victims and stabilize dangerous structures. They then carefully remove debris and structural elements. To prevent further collapse while ensuring safety of workers and victims, the rescue effort is by necessity planned and performed cautiously often taking several days. This delay in actual rescue is not without consequences. Figure 1. Canine equipped with CRDS and underdog When a US&R dog finds a live victim, it provides a bark indication which continues for some time as the dogs remains where it found the victim. The canine handler then uses a remote control to wirelessly release the underdog from the dog. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. "IWCMC 10, June 28 July 2, 2010, Caen, France. Copyright 2010 ACM /10/06/...$5.00"

2 Figure 2. CRDS detail. From left to right: Harness, CRDS with Bark Release, underdog. The CRDS has been tested by the Provincial Emergency Response Team (PERT) of the Ontario Provincial Police (OPP) in numerous US&R training exercises in Canada and the United States. The wireless transmitter release mechanism has been demonstrated to work effectively in most cases. There are however two conditions which may cause this version of the CRDS to fail. The first is when the dog goes outside the audio range of its handler. If a handler cannot hear the dog s barks, they do not know when to release the underdog. This situation may occur when the dog moves deep into inaccessible rubble where structural elements dampen the sounds of barking. The second condition is when the dog goes out of wireless range of the transmitter used to release the underdog. This includes cases where, for example, debris from a collapsed building may act as a barrier to wireless signals [4]. In addition to these problems, we observed that individuals who are designated to use the transmitter to release the bag often find the task very stressful. The reason being because, if they cannot see the dog, they do not know if their attempt to release the underdog was actually successful until the dog returns. In fact, a typical means of using the CRDS is for the handler to use an assistant to press the button on the transmitter as the handlers generally report that they are stressed enough simply controlling the searching dog. In order to resolve these problems, a bark detection system and deployment algorithm was devised for the CRDS forming a bark release system. Such a bark release system permits the deployment of the underdog without intervention from the handler in cases where the use of the transmitter system is undesirable or impractical. The rest of this paper describes the hardware and algorithm of the bark detection mechanism of the CRDS, the experiments conducted to validate its effectiveness and the discussion of those results. Finally we will discuss the future work related to this system. 2. BARK DETECTION With the existence of commercial products such as anti-bark training collars, the development time was considerably shortened. By utilizing the already efficient electronic circuit in an anti-bark collar, detecting a single bark can be easily achieved. In order to filter environmental noise, the audio sensor is placed so that it is pressed against the dog s throat rather than being positioned outwards where it could receive spurious barking from other dogs and possibly misleading ambient noise. More importantly the sensor does not measure the audio signal of a bark but instead measure the vibration created through the dog s body when it barks. Thus, the sensor only picks up signals related to barking from the dog wearing the collar. However, a single bark is not a clear indication that the US&R dog has found a victim since all dogs sometimes bark randomly and often make other noises such as whining, whimpering or growling. The difference between US&R dogs and other dogs is that the US&R dogs are trained and rewarded for barking continuously when they have located the area of strongest human scent which usually gives a very good indication of the vicinity of the human victim. To filter out false indications, an algorithm was developed that differentiates between single and multiple barks and between multiple barks in quick succession and random barks over long periods of time. Figure 3. Bark Release Algorithm After the first bark, a timer is started and will continue to run for 8 seconds. In the duration of 8 seconds if 4 more barks have not been detected then it reverts back to the beginning state where it waits for another first bark again. If 4 barks are finally detected within the 8 seconds then a delay of 5 seconds is introduced followed by a signal to the CRDS to release the underdog. The 5 seconds delay was introduced as it was observed that the dogs would either stay where they are after barking or they actually try and move closer to the source of the scent. In either case, five seconds does not harm and in the previous case it allows the underdog to fall closer to the victim. The algorithm is depicted in Figure 3. The algorithm was implemented on a microcontroller and introduced to the CRDS. The electric circuitry of an anti-bark

3 collar was removed from the unit. In its originally intended use, the microphone circuit output was to drive a solenoid valve that was intended to spray a harmless but annoying air or chemical to deter the dog from barking. The circuit was modified so that its output was redirected to the microcontroller. The microcontroller s output was connected to the original CRDS release mechanism. The new CRDS could now be manually activated via its original transmitter or automatically activated through the bark detection system. 3. EXPERIMENTS Two sets of experiments were conducted to ensure the effectiveness and accuracy of the bark release CRDS. The first experiment was a proof of concept test and the second was designed to determine the expected accuracy of the bark release CRDS. A typical exercise in the training of US&R dogs is called the bark barrel. A person to be found by the dog (called a quarry ) is placed inside one of three large barrels and hidden from view as shown in Figure 4. by carrying the equipment or having a piece of it mysteriously fall off. The second experiment was design to ensure that the automatic bark release algorithm can deliver the underdog as close to the victim as possible employing our algorithm. The original CRDS performed fairly well in previous tests with a human monitoring the barks and manually triggering the release. This experiment was setup to compare the performance of the manual release and automatic bark release. A quarry was placed in a rubble pile and concealed in a hole beneath a wooden pallet. The centre of the wooden pallet was marked with orange spray paint as shown in Figure 5 as a visual guide to the human observers on the rubble pile. Figure 4. Bark Barrel Test The purpose of the bark barrel is to get a dog to bark when it smells the human in the barrel. Eventually, the barrel is moved to a rubble pile and then removed altogether as the dog becomes used to finding hidden humans in voids and hidden spaces in rubble piles. The bark barrel was used as the basis of our first set of tests as the barrel is a fundamental part of all US&R canine training and a good place to introduce the new equipment to the dogs for the first time. With the help of the canine handlers, the bark release version CRDS and underdog were worn by 4 different dogs while they performed the bark barrel exercise. The bark barrel exercises provided an opportunity to calibrate the CRDS to the barking characteristics of the dogs. This included determining the maximum number of barks that can reasonably expected when a dog indicates the presence of a hidden human MaxBarks, the determination of a reasonable interval within which that series of barks should be counted Interval, and the delay to allow the dog time for finally positioning its body before the underdog is released DelayTime. The tests also provided an opportunity to properly fit the CRDS harness to the dogs and to ensure that each dog was not distressed Figure 5. Location on rubble pile where the quarry hid Three dogs were sent out, one at a time, to find the quarry. A human assistant was provided to the canine handler who was not able to see the dog searching or the pallet target. The job of the assistant was to listen for the dog barking and trigger the release of the underdog when they thought the dog might close to the target. The distance between the centre of the target and the underdog was measured. Figure 6. Dog finds victim Following our manual release trial, the quarry was moved to a similar but different location on the rubble pile. This was done because we were using the same dogs for the second trial and the dogs are smart and quickly learn the location of the quarry from

4 the first trial and will run straight to them without actually searching in the second. Similar to the previous trial, each of three dogs were sent to locate the victim except the dogs were equipped with the bark release CRDS. Again, each time, the distance between the centre of the pallet and the underdog was measured. 4. RESULTS The bark barrel experiment was successful. The underdog was released at the appropriate moment (5 seconds after the 5th bark) on every test in all cases with all the dogs. The algorithm performed well under the different barking habits of the dogs. Some of them were excited and barked rapidly while others barked at a slower rate. The algorithm performed correctly in both cases. In general, the dogs adapted to the unfamiliar harness and equipment quite well. One of the dogs had never worn the harness before and the handler had never seen the equipment. Yet the team was able to deliver the underdog successfully without any observable hesitation or distraction on the part of the dog. The results of the accuracy experiments are presented in Table 1. Manual Release Automatic Relese Table1. Distance from underdog to victim 1 st Dog 2 nd Dog 3 rd Dog 90cm 122cm 60cm 75cm 75cm 360cm The average distance from the victim to underdog in the manual release trial is 90.67cm while in the automatic release trial, the average distance is 170cm. 5. DISCUSSION An important impetus for conducting these tests was to confirm that our equipment does not alter the search characteristics of US&R dogs. This is important because dogs do this task very well and, because they are so fast, they save lives by finding people in rubble very quickly. Our tests confirmed that our equipment did not hinder the dogs and provides them a new capability automatically delivering supplies to victims. This new functionality is significant in itself for a number of reasons. Firstly, this marks the transition of US&R dogs from a role of search to one of assisting in the rescue of humans. As it is unlikely that any artificial system is likely to move through rubble as quickly as a dog, this functionality makes dogs even more necessary for US&R work. Secondly, the dog now becomes part of the rescue management decision making process. It is the dog that decides where the trapped human is, makes the first contact and takes the first step toward rescue. In effect, the dog is an agent assisting humans effectively using Biological Intelligence (BI). It is also important that we discuss an apparent anomaly in the presented data. It should be noted that the search dogs find their way to hidden humans using scent. They search for human scent and follow the scent trail until they find the highest concentration within the scent plume. This is when they give their bark indication. The scent plume can be drastically affected by wind and the configuration of the rubble. On casual examination of the data, it appears that the manual release of the underdog appears to result in greater accuracy based on the longer distances from the target using automatic bark release. Drawing this conclusion would be wrong. The hiding locations for quarries called hides, are selected primarily for the ability of the quarry to actually occupy the position and remain hidden. The location of the first quarry was selected based on a convenient hole being near the centre of the rubble pile. There was no wind and the space occupied by the quarry was a culvert, sealed off from other parts of the rubble. Essentially, this meant that the dogs were not presented much of a challenge as the scent plume was emanating from a single location and was concentrated around the quarry s location. Figure 7. Circle indicates hide location, Square show pooling of smoke and, presumably, scent from the quarry. At the second hide the conditions were considerably different from the first. The second hide was located on a small hill formed by the rubble and the quarry was hidden in a hole that extended into the rubble pile and resurfaced at several locations around the pile. In addition there was a brisk wind blowing across the pile that entered the hide and pushed scent through the holes in the pile. As the dogs were taking much longer to find the second hide we investigated why the dogs were having such a hard time finding the quarry. After the testing, a smoke grenade was thrown into the hide and the dissipation of the smoke was observed. We discovered that the smoke was being discharged through at least 4 separate holes around the quarry s former location. In addition, the wind pushed the smoke (and presumably the human scent) over the hill and pooled it in a shallow bowl-like structure where one of the dogs eventually gave its bark indication and released its underdog. This situation is shown in the figure 7. The test was considered valid by the canine handlers, as in their work environment, unpredictable wind conditions are a fact of life. We present the argument that, if placed in the same situation, the human handler would still choose to release the underdog at the same place upon hearing the dog s barks. After reviewing the results of the experiments performed it can be concluded that the automatic bark detection feature is reliable and performs reliably with at least the same accuracy as manual release. At the same time, it works for situations where manual release would fail. 6. FUTURE WORK The success of the automatic bark detection release feature of the CRDS allows the exploration other research areas related to real

5 US&R problems. While others are researching methods of creating a network that is suitable for a disaster environment [5], we are exploring the possibility of using our research to deploy the network. We hope to use dogs to deliver wireless mesh nodes (radio repeaters in US&R parlance) to the victim and building a trail of wireless node breadcrumbs from the victim. Thus creating an ad-hoc wireless mesh network from deep inside the collapsed structure to the outer perimeter. Once the network is created, we hope to assist in the deployment of other systems to help with the rescue effort such as the Canine Augmentation Technology [6] and, potentially, rescue robots once a reliable network is available. Another area being explored is the design of miniature robots that the dog can carry and drop by the victim. Although locomotion may be limited at least the robot can perform reconnaissance tasks, giving a clearer views of the condition of inside the rubble and the victim. 7. ACKNOWLEDGMENT Special thanks go out to Constables Kevin Barnum, Mike Dellaire and Mike Dolderman of the Provincial Emergency Response Team, Ontario Provincial Police. 8. REFERENCE [1] Collins, L. Assessing structural collapse from acts of terror. City, [2] Hnatko, B., Key concepts & tactics in responding to a structural collapse, in Fire Rescue Magazine 25, 2007, pp [3] Ferworn, A., Ostrom, D., Barnum, K., Dallaire, M., Harkness, D. and Dolderman, M. Canine Remote Deployment System for Urban Search and Rescue. Journal of Homeland Security and Emergency Management, 5, ), pp [4] Yamazaki, K., Ito, M. and Watanabe, A. Reconstructing a Communication Infrastructure in a Time of Disaster. City, [5] Gao, M., F., Z. and Tian, J. Wireless mesh network for emergency response system based on embedded system. In Proceedings of the The 2008 International Conference on Embedded Software and Systems Symposia (ICESS Symposia 2008), July [6] Tran, J., Ferworn, A., Ribeiro, C. and Denko, M. Enhancing canine disaster search. In Proceedings of the IEEE International Conference of Systems of Systems (SoSE 08), Monterey, CA, USA, 2008.