It has been shown that trail cameras are much better at detecting predators compared to tracking tunnels, which are the default detection tool. Trail cameras are between 2-10 times more sensitive at detecting predators than tracking tunnels depending on species. This is obviously very important as we need to be able to measure all predators that are out there.
The goal of our heat camera was to create an even more sensitive detection tool that can also automatically detect predators using machine learning. The heat camera is more sensitive for three reasons
- There is no delay in the start of recording. It is constantly recording and saves whenever it detects a predator.
- The motion detection is done by software, so sensitivity can be adjusted. False positives are not as big an issue when machine learning will be doing the sorting.
- It is suspected that the heat detection device can record over a longer distance than trail cameras.
- The trail cameras used for the comparison are Browning (Down Ops Elite) and Little Acorns (5210a) based on a recommendation by our partners ZIP (link) who have done a review of a large number of trail cameras.
To measure the sensitivity we used three methods.
- The simplest method was marking out a distance and walking towards both the heat camera and trail camera to see when they detected something large like a human.
- Next, we put a trail camera beside the thermal camera and just compared the number of counts of different animals recorded.
- We also integrated the output from a trail camera’s motion detector sensor output directly into one of the computers which drives our camera. This allowed us to make systematic comparisons of motion detected by our camera vs. the trail camera
Heat cameras can see three times the distance of a standard trail camera
The results of the first test for large objects like a human were that, with the trail cameras, you needed to be within about eight meters before it would trigger. Whereas the heat camera was triggered from 24 meters. Our camera will be mounted on a motorised rotating head which means our camera could be able to detect over nine times the area of that of a standard trail camera.
Sensitivity is different for depending on the animal
Obviously, the camera will be more sensitive to some animals than others. The table below shows how often the trail camera detected animals compared to the Cacophony camera. This is just a very small sample from one spot over six nights. The goal of this blog post is to give a quick indication of how much more sensitive the heat camera is, rather than a rigorous analysis. That said, the initial results look quite encouraging!
Trail camera counts
Cacophony camera counts
Times more sensitive
What is very clear is that we can detect a lot more with the heat camera. I suspect with a larger sample size bird detection will be shown to be even more sensitive. We were quite surprised how many of the possums were missed by the trail camera. It seems they were missed from either being too far away for the other camera to detect, or they walked past too fast for it to turn on.
This photo shows the camera set up. The red box at the bottom is where the heat camera is. The fine human specimen is just there to give perspective. To the right is a live capture trap that was baited with meat along with a tracking tunnel. On opposite side of the river, there are another tracking tunnel and chew cards.
Below is an example of a video from one place over six days that shows how each of the predators look. It is a bit worrying that this one random spot could see so many different predators in just one week!
Video 1: Example footage from heat camera - predator everywhere!
Seeing so many birds is probably a distraction if you have to look through the footage manually to find predators, but the goal of this is to do the detection automatically. It seems like this will be a fantastic bird detection tool as well. Imagine how easy it would be to work out the efficacy of something like a poison drop if you could measure the background birds and predators for a week before and after the treatment.
Even though this device will cost more than a standard trail camera, the increased sensitivity and dramatically lower labour costs from machine learning mean it will be a vastly cheaper and more accurate way to be able to monitor the environment. It is all digital so the costs will plummet over time. It shouldn’t really come as a surprise that a device designed for hunters tracking deer and pigs is not that great for spotting the small mammals we are interested in.
Imagine if you were studying the stars and had a telescope where you can see five times as much....it has to help. This does not mean we have 100% detection because until there is a more sensitive tool we won’t know what is missed. The same hardware and software will be able to be used on a higher resolution heat camera that is the most likely way to further increase sensitivity.
The biggest drawback with the current camera is the power requirements are high because the camera is always on and it is running a small computer. When we have worked out our requirements for recording and running the machine learning model there will then be some well-understood work to be done to optimise the hardware for lower power consumption.
Our main short-term goal is to get enough videos of our target predators so we can teach the machine learning (artificial intelligence) to automatically recognise different predators. We currently have about a thousand tagged videos for training the machine learning model. More on our initial progress with machine learning soon…