Animal Detection when Game Camera is Moving

[Explorer]

The Rowcliffe paper (see source below), provided a formula for estimating animal detection using fixed game cameras.

I wanted to estimate how animal detection will change if the game camera is also moving.

A moving game camera is equivalent to human eyeballs as detectors either hiking thru a forest or driving in car.

The authors developed a theoretical formula linking the animal density to the photo capture rate using properties of the camera.

The expected number of contacts between animals and camera traps (Y) (each contact giving rise to one photograph) was estimated by the following formula:

Y = (2 + Θ )/Π * rTV * D

 

Where,

T = time

V = animal speed of movement

D = animal density  (number per square kilometers)

r = camera radius of detection

θ = camera angle of detection

 

 

If we convert time and animal speed into the range covered by the animal in one day, then number of game camera contacts per day is proportional to the range x density. 

The question, however, is what happens when the game camera is also moving (equivalent to a hiker taking a 20 mile hike into the forest with GoPro camera) sunrise to sunset as an example.

Would the probability of detection increase, decrease, or stay the same because the camera is moving?

 

I have not worked out the math and don’t know the answer, but my initial impression is that the probability of detection would increase.  I think this explains why detection of BF while driving yields so many reports.  Multiple cars along a highway provide continuous game cameras (24-7) via the eyeballs of the many drivers along a larger distance than just a fixed point in the highway. 

 

Although, the correct comparison would be just one driver going back and forth (for example) a 40 mile stretch of road 24 hours a day vs. a fixed camera on the road.  Assuming that the BF road crossing is random and not predictable, which method is more likely to detect the BF crossing the road?

The tricky thing about this problem is that the area covered by the camera (whether on game camera or on car/hiker) is the same.  So we are not getting more visual coverage in the camera lens.  The only difference is that the detector is moving. 

 

This problem has probably been answered by the US Navy when looking for German subs.  Which way is better to detect an enemy sub, by the ship staying put or moving in a zig zag fashion or other?

 

My gut feeling is that the increase in detection would be a multiplier that is a function of the additional range covered by the moving camera.  Maybe raised to a fractional power? 

I need to do more research on this.

 

Any suggestions?

 

 

Source:

 “Estimating Animal Density using Camera Traps without the Need for Individual Recognition” by Rowcliffe, Field, Turvey and Carbone in Journal of Applied Ecology, 2008.  See link below.

http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2664.2008.01473.x/pdf

[bipedalist]

Wire it up like the NFL maybe?!

 

Including sound dishes, etc.

 

Ziplines without cameras are for kids!

Edited September 22, 2018 by bipedalist

[Catmandoo]

They are in England. The paper is from 2008. The work was done in an enclosed animal park = zoo. Their position is that 'individual recognition' is not needed. I have downloaded the paper and will look at it in more depth. On this continent, I believe that animal researchers have at times,  gone back to 'white flash' on their cameras to identify animals better by spots, coat color and stripes etc.. They did not like the IR flash images. Last time that I looked at the Spypoint website, there was a presentation on their 'antler recognition' software. Haven't followed up on that.

We understand about the cost in money and time. Still boils down to boots on the ground. It is difficult to get an accurate count while we continue to scare them away.

[gigantor]

1 hour ago, Explorer said:

 

The expected number of contacts between animals and camera traps (Y) (each contact giving rise to one photograph) was estimated by the following formula:

 

 

Y = (2 + Θ )/Π * rTV * D

 

Where,

 

T = time

V = animal speed of movement

D = animal density  (number per square kilometers)

r = camera radius of detection

θ = camera angle of detection

 

If we convert time and animal speed into the range covered by the animal in one day, then number of game camera contacts per day is proportional to the range x density. 

The question, however, is what happens when the game camera is also moving (equivalent to a hiker taking a 20 mile hike into the forest with GoPro camera) sunrise to sunset as an example.

 

Would the probability of detection increase, decrease, or stay the same because the camera is moving?

 

 

Does r mean: the radius where the camera sensor gets triggered?  or just the theoretical area of coverage?

 

I think when the camera is moving, r would decrease or would have to be adjusted so as to not produce many false positives.

[Explorer]

10 hours ago, Catmandoo said:

They are in England. The paper is from 2008. The work was done in an enclosed animal park = zoo. Their position is that 'individual recognition' is not needed. I have downloaded the paper and will look at it in more depth. On this continent, I believe that animal researchers have at times,  gone back to 'white flash' on their cameras to identify animals better by spots, coat color and stripes etc.. They did not like the IR flash images. Last time that I looked at the Spypoint website, there was a presentation on their 'antler recognition' software. Haven't followed up on that.

We understand about the cost in money and time. Still boils down to boots on the ground. It is difficult to get an accurate count while we continue to scare them away.

 

Hi Catmandoo,

 

Yes the paper is theoretical and based on fundamental physics of a generic game camera and movements of the animals.

My recollection is that the authors used the physics analogy of a particle in the box to generate the primary equations.

The field test of the equations was limited, I agree, but necessary to test if they yielded reasonable estimates.

I have not found any better papers in the internet addressing this question (how many cameras are required to detect an animal given its range and density).

Also, I have not found any papers addressing my new question, what happens if the camera is also moving. Which is completely theoretical, since I am not proposing to have game cameras moving in the forest.

I am just trying to understand if hikers and cars (who ar moving and have eyes for detection) increase or decrease the odds of detection vs. a fixed camera (given the same amount of time in the field).

Key assumption here is that human eyes are as good as IR cameras for detection of movement.

 

Thanks for your interest and please let me know if you find anything that could help me better understand this.

 

 

10 hours ago, gigantor said:

 

Does r mean: the radius where the camera sensor gets triggered?  or just the theoretical area of coverage?

 

I think when the camera is moving, r would decrease or would have to be adjusted so as to not produce many false positives.

 

 

Gigantor,  

 

My understanding is that r is the distance between the camera and animal when it was detected by the IR sensor.

Below is a quote from their paper on the method used to estimate r.

 

The variables were estimated by a series of trials in which the camera was approached by a person at varying speeds and from varying directions, recording whether and at what point the sensor was triggered. In order to estimate the
detection distance, r, , the zone was crossed perpendicular to the sensor beam.

 

I suspect that the radius of detection will depend on specific camera design given the large diversity and selections available.

Nonetheless, I am more interested on the macro relationships between detection and range and density and not so much on the "constant" variables in the equation.

 

I agree with you that r will be very different for a hiker or driver in car using a different camera or just their eyeballs.

 

BTW, my question is completely theoretical since I don't think you could place a IR triggered game camera in a car and drive without it triggering it all the time with the car motion.

My interest is more with the analogy to human eyeballs.

Maybe human eyeballs are even better than IR detection given our evolutionary history to pay attention to movement in the bush?

 

Edited September 22, 2018 by Explorer

[scottv]

Part of the equation factors in the movement of the animals. If an animal moves a lot or has a large home range they are easier to detect. Other papers that have used this method have gotten the movement rate from a previously published telemetry study. This method estimates the density. Not just presence absence or how many cameras placed for what time period to detect said animal. As stated above one problem is camera placement. The paper is a little vague on this and there have been some back an forth "discussion" in journals on it. The cameras sort of have to be randomly placed but still need to have a good chance at detection. If they're all on used game trails then that biases the data, however you can't have cameras facing a rock wall either. 

 

There's a been a bit of work to come up a with a method to estimate density with trail cameras for animals that don't have readily identifiable individuals. There's been a few papers using this method.(tapirs, a southeast Asian pig, Iromote cat in japan). Interesting idea to try to apply the method to human eyes. Hmmmmm....would you have to quantify things like changes in direction where you're looking? How many head turns per hour?

[Catmandoo]

10 hours ago, Explorer said:

BTW, my question is completely theoretical since I don't think you could place a IR triggered game camera in a car and drive without it triggering it all the time with the car motion.

My interest is more with the analogy to human eyeballs.

Maybe human eyeballs are even better than IR detection given our evolutionary history to pay attention to movement in the bush?

 

In theory, you have to exclude some sets of human eyeball groups like baseball umpires as they relate to motion detection.

Camera PIR sensors tend to have a narrow field of view, short range. Wind moving brush causes false positives on a fixed camera. Moving PIR is questionable. Animal population studies have been carried out with fixed mount Trailmaster units. The passive unit has an area of sensitivity  of 150 degrees, range about 65'.

Human eyeballs suck at night. Rod vision is limiting. It is possible to set up some cameras with motion detection at the viewfinder. One of the basic auto-focus methods is 'contrast detection'.  Accessories are available for limited brands to display motion on smartphones. There may be an app that sends input to a motorized pan & tilt, to track the object in motion. I am waiting for an email from the manufacturer. Night work would require supplemental IR lighting.

Sooooooooooo, Explorer, theoretically speaking; contrast detection, now more than ever.

[Explorer]

My suspicion was correct; there is a lot of literature on mathematical models of search and rescue by humans (using eyeballs) by US Coast Guard, US Navy and academia. 

Rather than re-invent the wheel, I am just going to borrow one of these models to explore the question (does success at detecting BF greater with vehicles than with fixed cameras in a large fixed area).

One good summary of the history of search models used by USCG is the report titled:

“Review of Search Theory: Advances and Applications to Search and Rescue Decision Support”, published in 2001

See link below:

 

https://www.researchgate.net/publication/235193662_Review_of_Search_Theory_Advances_and_Applications_to_Search_and_Rescue_Decision_Support

 

The search model equations are simple.

POS = Probability of Success = POC x POD

POC = Probability of Containment, is the likelihood that the target is contained in the area being searched.

POD = Probability of Detection within the area being searched.  It is usually a function of the search effort density (or coverage, C)

The probability of Detection will be estimated as: P (t) = 1 – exp(-Wvt/A)

 

The fraction Wvt / A is the density of search effort in the region. In operational search and rescue terminology, this fraction is called coverage.

The numerator Wvt is called search effort or area effectively swept (Z). Thus coverage is the ratio of the area effectively swept to the amount of area searched.  This equation is called the random search formula.

 

Below is a summary of the Variables Definitions used in the search model:

Effective search (or sweep) width (W).  Distance on either side of road or trail that is searched by humans.

Effort (total length of the searchers’ tracks while searching, L = vt)

v = velocity of vehicle or hiker

t = time

Search Effort (area effectively swept, Z = W x L)

Search Effort Density (Coverage, C = Z/area searched)

 

In the example I will use to explore this model, I will assume there is a 100% certainty that there is a BF in the Olympic Peninsula in an area of 50 x 50 miles (2,500 square miles).

The target area to search for the BF is the full 2,500 square miles.

The BF location is unknown but its daily range will be assumed to be radius of 10 miles.  Thus, every day the BF could be found in an area of about 314 square miles within the 2,500 square miles of the National Forest lands.  This yields an estimate of the Probability of Containment POC) of 314/2,500 or about 0.13. 

 

The search effort will be done in one day via 3 mechanisms:

 

Vehicle Detection Mechanism:

•            One 40 mile road goes thru the NF and about 2,000 cars travel the full length of this road every day.
•         Assume average speed is 40 mph, so every one of the 2,000 cars spends one hour on this road.

•            The drivers of the cars are able to detect any animal moving 25 ft to the left and to the right of the center of the road ahead of them.  This will be the sweep distance assumption (W).

•            Total area searched in one day per car is about 0.189 square miles.

•            Total area searched in one day by 2,000 cars is about 379 square miles

•            Total Coverage = 379/2,500 = 0.15

 

Hiker Detection Mechanism:

•        There are hundreds of trails into the forest from the road.

•            Assume that 100 hikers go into the forest every day and hike an average of 10 miles.

•            The hikers are able to detect any animal moving 25 ft to the left and to the right of the center of the trail ahead of them.  This will be the sweep distance assumption (W).

•            Total area searched in one day per hiker is about 0.05 square miles.

•            Total area searched in one day by 100 hikers is about 4.7 square miles

•            Total Coverage = 4.7/2,500 = 0.0019

 

Fixed Game Camera Detection Mechanism:

•            There are 20 game cameras placed randomly in the forest.

•            Each game camera has an effective area of detection of 690 square feet.

•            Total area searched in one day for the 20 cameras is 0.0005 square miles.

•            Total Coverage = 0.0005/2,500 = 2 x 10^^-7

 

 

With these assumptions the resulting POD and POS for a one day search for each method is summarized in the table below.

 

 

Per these results, the probability of successfully seeing a BF with vehicles is low, but is about 74 times more likely than with hikers and ~700,000 more likely than with 20 fixed cameras.

Granted, these results are based on our simple assumptions.

Maybe a driver and a hiker are not good sensors for animal motion +/-25 ft from the road/trail.  Maybe their field of view is much narrower or they are just looking ahead of them and not paying attention to what is going on the sides.  A camera, on the other hand, will be detecting motion for 24 hours with its passive IR sensor.

 

Nonetheless, this simple model does suggest that vehicular traffic as sensors searching for BF motions within the target area will have a higher success rate (because of the larger area covered) than fixed cameras or hikers.

The estimated POS numbers are low as expected because of the low POC.

The following charts are sensitivity charts showing how the POD and POS changes with assumptions in number of vehicles, hikers or cameras.

 

 

[scottv]

The longer you leave trail cameras out the higher the detection probability. And if you have a population of animals (not just one) that move quite a bit, the probability goes higher. Because you don't need to check them very often once they're set up, they end up being a very efficient way to sample a habitat. Quite a few papers on this, but I think this is pretty good:  https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0175684&type=printable

[scottv]

Actually this one might be better for what you have in mind. Density estimation of 2 snake species from road crossing observations: http://www.dtic.mil/dtic/tr/fulltext/u2/1046415.pdf

 

 

[Explorer]

18 hours ago, scottv said:

The longer you leave trail cameras out the higher the detection probability. And if you have a population of animals (not just one) that move quite a bit, the probability goes higher. Because you don't need to check them very often once they're set up, they end up being a very efficient way to sample a habitat. Quite a few papers on this, but I think this is pretty good:  https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0175684&type=printable

 

Your summary makes perfect sense and is consistent with the findings in the Rowcliffe paper.

 

17 hours ago, scottv said:

Actually this one might be better for what you have in mind. Density estimation of 2 snake species from road crossing observations: http://www.dtic.mil/dtic/tr/fulltext/u2/1046415.pdf

 

 

I could not open the first link, but was able to open the 2nd one. 

Thanks for the paper reference!  I will certainly read it to learn how others are estimating density.

 

Edited September 25, 2018 by Explorer

[Explorer]

BTW, I think the search detection model that I used above is not very good because it does not reflect the value of area covered x time.

 

The POD (probability of detection) is only a function of area covered, so it is penalizing searchers that are slow or stay put in an area.

 

For example, if a car covers 40 miles at 10 mph, the observer in the car is spending 4 hours looking while the same car going 40 mph is spending only 1 hour looking.

From an area covered point of view, the area is the same - 40 miles times the sweep distance

But from an area covered x time, the slow car is 4 times more into search mode.

Time x Area covered must be valuable in search since the object being searched could be moving randomly, and there should be no penalty for searching the same area more than once.

 

The problem is magnified for a stationary camera where the value of a fixed area being observed 24 hours gets no credit.

Thus, the findings/observations using that simple model shown above are flawed.

 

I still think the best way to solve this problem is with simulation model and not with parametric models (like the Search model or Rowcliffe model).

In a simulation computer program you can have multiple agents (searchers and BF) behaving according to their nature (either stochastic or predictable or combination) and uncertainty can be determined by running the simulation thousands of times.

In the example above, I was just looking for a short-cut approximation.

 

 

Edited September 25, 2018 by Explorer

[scottv]

Try this one, trail camera study in Japan:  https://www.jstage.jst.go.jp/article/mammalstudy/29/1/29_1_37/_article

 

[scottv]

For the link that you couldn't open, try entering this into google: "Camera trap arrays improve detection probability of wildlife: investigating study design considerations using an empirical dataset".

[scottv]

I think I posted the link for this project on another thread here, but this is a wonderful long term trail camera project run by volunteers. The yearly reports are very good but don't think they have the statistical detail you're looking for. However you yourself might be able to do something with the data.

http://www.conservationnw.org/our-work/wildlife/wildlife-monitoring/