[Music] and we're right on time and for our last um talk which is going to be our longer talk and that is from Akos Liti and he's going to be talking about his work developing an animal-born acoustic shockwave detector so Akos whenever you're ready take it away thank you thanks for the introduction uh I used to be at Hungary but now I'm at Vanderbilt University in Nashville Tennessee in the USA and this work has been a collaboration with lots of people and and uh institutes as you can see on the screen uh to make sure that uh I have time for the acknowledgement i put this slide in the beginning so this uh project has been funded for by Vodafone and also Vanderbilt University and uh these wonderful institutes institutions at the bottom have been helping us to to actually deploy and test uh the prototypes all right so what what was our motivation to to to go into this project well of course we all know that poaching is a problem so anything we can do to help with that would be helpful and my group at Vanderbilt has been working on eostic gunshot detection classification and localization for over 20 years we have been using local sensor specifically wireless sensor networks so uh that was kind of uh uh one of the motivations and we have heard about static deployments of shot detectors uh to protect wildlife but the problem with those that they are not scalable they protect a relatively small area and not necessarily the animals themselves and then of course we have witnessed the proliferation of GPS tracking colors so out of these four things came the idea of let's create a shot detector that the animals themselves uh wear so it protects the animal so let's talk a little bit about gunshot acoustics so typically when people hear a gunshot they hear the muzzle blast which is basically the the propellant of the bullet exploding in in the barrel but if it's a rifle shot then more than likely it is a supersonic uh projectile that is leaving that barrel and the supersonic projectile generates a sonic boom also called a shock wave and that can also be detected by the microphone so if you look at uh the typical acoustic events when you when you record a gunshot the first one is typically the the the shock wave itself and the beginning it really looks like an N a capital N and it's a very unique event then you hear of course ground and other reflections and then eventually the muzzle blast uh arrives at the microphone as well the shock wave is such a unique event uh that it's easier to detect and and classify as a shock wave so it's a better uh uh signal to focus on so just briefly what what you can do uh with microphones when when a a shot is being fired so here S denotes the gun that's being fired and then you can see the trajectory of the supersonic bullet and then M denotes the sensor so if you have a microphone there you can detect the time of arrival of the muzzle blast the time of arrival of the shock wave and even the length of the shock wave which depends on the caliber and how far the projectile is from the sensor now if you have multiple of these sensors you can actually locate the shooter if each of these sensors have multiple microphones so a mini microphone array and we have been building let's say arrays that are only let's say three inches aside so four microphones then you can also detect the angle of arrival of all of these events basically the direction toward the shooter because the sound arrives at those microphones at different times and from those time differences you can figure out the angles now if you do all of these things and you have multiple of these sensors you can locate the shooter you can exactly identify the trajectory in 3D you can uh you can uh estimate the speed of the bullet the caliber of the bullet and then even what kind of rifle it fired it so that's just uh for for interest that with these low cost sensors you can do a lots of things with with gunshot echo sticks now of course for this project the only interesting uh thing is the presence of a shock wave so if if a bullet flies close to the sensor we want to able to detect it that hey there was a gunshot so that leads us to the overall concept of the system so we want to integrate a shot detector into a GPS color and when there a shot is detected a shock wave is detected we want to immediately send an alarm uh with the latest GPS uh uh tag that the device has we don't want to wait for the GPS uh acquisition time because the meantime they may destroy the sensors but of course we wake up the GPS it um gets a new location so we can send an alarm uh with a better position than than originally uh we had that information all right so again this slide just shows that we don't want to reinvent the wheel so this unit is an add-on to existing GPS uh commercial GPS colors so optionally they'll be able to integrate it into their uh system and so our work only focuses on the board that detects the the the shot and then communicates this to the GPS colors so this project started in 2007 a long time ago so this is our prototype from 2008 uh the this was for elephants specifically so that's why this is kind of big um one kind of important factor that we actually used two different microphones so a PSO microphone is attached to the wall of the sensor and it's great to detect mechanical impacts a mechanical impact may look like on the microphone similar to a shock wave but with the PSO pickup and the regular microphone having those two different signals we can actually fairly well distinguish between uh mechanical impacts and actual shock waves and then also there is an accelerometer on that on that board so we can also monitor the change in the animal's behavior after a suppose shot we should see a change if it was a shot if it wasn't a shot then probably there wouldn't be any change now the prototype also had an SD card just to record these shots for further refinement of our detection algorithm of course the final uh device will not have uh the SD card so we tested the prototypes in various settings on cattle here in in Tennessee and the San Diego Zoo was nice enough so that we were able to deploy it on on captive elephants for a couple of weeks we had some uh deployments in Kenya and additional tests so we gathered lots of uh data to be able to test the detection algorithm this is kind of summarizing uh the findings oops so uh you see on the scale the detection output is uh probability from 0 to one 0 to 100% of whether this was a shot or not and you can see that even this prototype from a few years ago was was pretty successful in in distinguishing uh gunshots from other natural noises specifically animal sounds or mechanical impacts and and uh other nature sounds so that this gives us hope that this will uh work well again this is a work in progress so we don't have a final product ready to be uh deployed all right this is just uh summarizing various ways of of detecting poaching specifically with guns and uh basically I think our approach kind of hits the sweet spot so for example it doesn't only uh uh detect I mean does protect one animal uh the shock wave can be detected reliably within 50 mters uh so if there is a herd situation then a single sensor can can potentially detect u protect the entire herd now when I say protect of course there is a caveat which is of course if somebody is already shooting at those animals it's probably too late for that poor animal but if the alert goes out and the rangers can come and and uh catch the poachers then of course it really helps the the next animal they they would have killed all right so summary so I described this variable gunshot uh shockwave detector idea so it's still a work in progress of course it has to be low power and we did uh various uh tricks to to make sure that this uh can last for about a year depending on lots of factors of course uh I mentioned that we are using the vibration and sound uh to make sure that the false detection rate is very minimal because of course you don't want to call the rangers if there is no shot because eventually they will lose uh uh interest in such a system all right so we are still working on this we need to finish uh the new hardware design uh update the detection algorithm with modern AI based techniques to make it even better we still need to implement this motion based uh behavior classification after uh a detected event and actually we we are working with Sava tracking out of Kenya to to integrate this into their color at first and of course we need to test it uh I'd like to make sure that everybody knows that this will be open source and available for everybody so it's it's uh even though we work with a commercial vendor the the design and the software will be released so anybody can integrate into their uh devices if they want to all right so that's that was the short project but I briefly want to talk about another project that is ongoing so this is a national science foundation uh sponsor project started in 2023 so the goal here is to related to of course the shot uh detection but uh build an intelligent acoustic blogger so a low power animal born acoustic uh data gathering device so here you can see our current prototype as you can see it's very tiny it's hardly larger than an SD card and it has various operating modes of course typically people have been using schedule based so you can just deploy it and specify when it should record and when it shouldn't uh you also have a threshold trigger so when it's completely silent it's not going to just keep recording uh uh nothing but more interestingly there is this AI based filtering and clustering method that we are working on basically we are trying to identify events of interest and uh interesting events and rare events and not fill up the the SD card with detecting the same kind of noise over and over again so this is an ultra low power design so it's uh power consumption is lower than micro moth but it has much more processing power which is actually nicely tunable depending on what you want to do on the on the processing side and of course uh a version of this will be used in the shot detector because it's it's much smaller and much lower bar than the first prototype all right so this is based on the ambigo 4 MCU it's an it's an ARM core again we are shooting for continuous mon monitoring for up to a year but it depends on a lot of factors your on your battery size or your sampling rate configuration settings uh SD card size uh well the goals of the project and and so on so forth so the AI based approach is an unsupervised clustering algorithm that will create these these clusters of similar sounds and based on the deployment we can we can either record all those sounds or just keep statistics about the clusters and and representative samples out of those clusters you can also add an additional classification uh layer where you can specify whether you want to reject certain sounds so for example we can classify wind noise and not record continuous wind noise or if if you are specifically interested in a certain animal call then we can have a classifier that makes sure that all of those kinds of sounds are recorded and and uh the clustering runs on the rest of the the events all right so I think I'm going to wrap up here uh because time is running out i just want to show some ballpark uh power numbers on this uh last figure uh if you are interested thank you so much uh any questions yes we've got first of all thank you for that talk that's super super interesting um and we love to hear that that shockwave detector project is open source um because we love that and like I think Frank mentioned in the chat we're going to be having a a meeting on on funding and finance open source projects soon so we'd love to hear about that um we've got a lot of questions in the chat and some time to go through them so I'm going to pick a few out to um just read um there's one from from Nathan Harris that's a question on the AI based filtering that you mentioned um asking about what AI models they use um because that's it's interesting to be able to provide AI based filtering um based on the size of the device okay so so it's not one of the traditional models for for classification it's specifically an unsupervised clustering algorithm uh so it creates a lower dimensional space where it it groups similar uh sounds together and the part where we'll decide what to record and and what to discard is still not implemented but the clustering itself works very well and uh and again the goal is to somehow let the domain scientists decide what they want to record and uh if we have a a cluster with thousands of events we probably don't want to fill up the SD card and use all our our power on that so in that case we'll just let's say K store the metadata for those events and just some representative samples but the the goal is to provide a configuration interface before deployment time to specify all these requirements and automatically configure the software for that uh deployment great thank you um and then we have two questions from Cari one about which we can address in the chat if there's like a website or contact information for Wiper and also what the cost of the whole system is and Carly feel free to jump in i'm not sure if you meant the the um shock wave detecting system or that second project that a mentioned so so I mean the cost will be similar uh and and I assume by cost you mean how much it takes uh or cost to manufacture one of those boards uh I don't have exact numbers because right now we are just running u prototype development and of course manufacturing five boards is almost the same cost as 200 uh so right now I would say that the prototypes let's say a couple of hundred but if we go up to a few hundred uh batch of few hundred it's probably going to be under $100 in in hardware cost amazing and then there is another question um from Nathan that I think was like a clarifying question um that that you mentioned that the microphones on the on the um shock detector collars microphones are only four inches apart um and can and can calculate trajectory angle no no no no so so that that whole multimicrophone configuration uh was for just a general overview of what kind of things you can do with with with with acoustic detection of guns shots so this uh you the the shot detector only has a single microphone and the PSO just for for helping with with classification so the only thing we can we can detect with that or with this unit uh is there was a shot period okay no location or anything else other than the the sensor's location from the GPS scholar got it yeah that's it's a super fascinating project and seems like it has a lot of really um incredible potential use cases um and I know we're we're right on time so thank you so much for that for that great long talk um feel free to keep putting questions in the in the chat for those who are listening um and then Akash can can hopefully um pop in there to answer them um so again thanks everyone for joining our variety hour this month and for our speakers um for their lovely presentations i'm going to end the recording in just a few seconds but everybody is invited as always to stick around for after hours um to have a more casual conversation about um anything that we heard today or really any other topic you could think of um so thanks for coming and we will see you next month on April 30th bye [Music]