Trial results: AI heat camera is more sensitive and cheaper than trail cameras for possum detection

AI heat camera is 3.5 times more sensitive than trail cameras for possum detection: over 200 times faster to analyse and almost 8x cheaper per detection

James Ross - Lincoln University and Grant Ryan - The Cacophony Project

This report outlines the initial test results from a trial funded by OSPRI to test the sensitivity and cost of different possum detection methods.

An experiment was set up in a two hectare predator enclosure managed by ZIP. A Cacophony Project's Artificial Intelligence (AI) heat camera was placed alongside a trail camera and chew card in three different locations in the enclosure. Over the course of a month, a possum was released for three nights on eight different occasions.

The following table shows the details of the possums released during the trial:

Trial #Animal IDSexWeight (kg)Date InDate Out
Heat camera & trail camera setup

The video below gives an example of the different types of video from the trail camera and the AI heat video. The heat videos use AI to identify where the animal of interest is in the video and then attempts to classify the type of animal. The label and number at the top is the best guess and the guess at each point in time is shown at the bottom. More often than not you can see that the camera identifies each detection correctly as a possum.

The video shows how the AI heat camera captures the full interaction of the possum coming in and out of frame but the trail camera only captures a small part of the interaction. This is because the trail camera has a delay between when motion is detected and recording starts, and the camera does not detect when the possum leaves. The first six videos show the heat camera view followed by the complete (much shorter) trail camera recording. This is followed by a number of possum encounters not captured at all by the trail cameras. Finally we show one video that was recorded by the trail camera and not picked up by the heat camera. This was a very close interaction where the possum may not have been in front of the heat camera.


The table below shows the detection times for the different devices. The chew cards were placed both 10m in front of the cameras and 20m in front of the cameras. Neither camera type was sensitive enough to detect a possum at 20m but none of the chew cards directly in front of the cameras were chewed. It is difficult to determine the relative sensitivity of the other chew cards as they may have had more interactions if there had been a camera close enough to see them. The chew cards were one of the only things of interest in this grassy enclosure (apart from food) so the interaction rate with them may well have been higher than what is seen in natural environments.

TimestampHeat CameraTrail cameraChew card - cameraChew card - other
16/11 10.02pmXX  
16/11 10.15pmX   
16/11 10.18pmX   
19/11 03.09amX   
21/11 10.18pmXX  
24/11 12.10amXX  
24/11 10.32pmX   
24/11 11.46pmX  X
25/11 12.29amX   
25/11 01.36amX   
27/11 09.38am X  
30/11 09.50pmX   
01/12 05.17amX   
02/12 02.34amX   
03/12 03.08amX  X
03/12 03.22amX   
07/12 02.10amXX X
12/12 09.58pmX   
12/12 10.27pmX   
Total detections18503

These results show that the AI heat cameras were 3.5 times more sensitive at detecting possums than the trail cameras. This data confirms a trial conducted in 2017 which showed that the AI heat camera was approximately 3 times more sensitive for possum detection. This is not surprising given that trail cameras are designed for larger game animals such as pigs and deer. Trail cameras can miss detections for two reasons: the motion sensor used to trigger recordings is not sufficiently sensitive and the camera has a delay in starting so misses fast moving animals.

While it is clear that the heat camera is more sensitive, the other notable difference between the camera types is the time it takes to analyse collected videos. During the trial period there was some particularly bad weather which resulted in a large number of false triggers creating many videos to be manually analysed.

The following table shows the number of recordings made and details around analysis of recordings:

 Trail CameraAI heat camera
Number of recordings5264342
Average recording length (seconds)1016
Number of recordings to be analysed526449
Minutes of recordings to analyse877.3313.07
Number of possum identifications518
Minutes of analysis per detection175.470.73

The AI heat camera produced vastly fewer false positive videos. More importantly, 297 of the videos were automatically tagged as “not containing animals” leaving only 49 videos to look through compared to 5264 for the trail cameras. It should be noted that the weather was quite windy during the trial period causing the trail cameras to generate a higher than normal number of false positives.

The AI camera's automated detection capabilities will continue to improve over time, further reducing the time taken required to analyse the recordings it generates.

Additional advantages of the AI heat camera

There is significant time involved in retrieving SD cards from the trail cameras, transferring them to a computer for analysis and storing them in an organised way. With the Cacophony Project's AI heat camera, videos are automatically uploaded to a central cloud-based database (if WiFi or 3G connectivity is available).

The heat camera also supports wireless walk-by data retrieval where there is no connectivity. A phone with a companion app can connect to the camera and retrieve videos which will then be uploaded when the phone has a connection. Wireless retrieval means the camera does not need to be opened to retrieve a SD card - a significant benefit in adverse weather.

Analysis of the cost of predator detection

When calculating the total cost of detection, the follow factors must be considered:

  • The cost of physical hardware
  • The cost of data retrieval
  • The cost of analysis

The table below shows an analysis of the cost per detection for the three different detection technologies tested.

In this experiment the time for data retrieval for the AI heat camera is actually 0 as recordings were automatically uploaded to the Cacophony Project API server using the 3G network but for this analysis we assume recordings are to be retrieved using the walk-by method described earlier. It takes more time to retrieve a SD card from a trail camera and transfer recordings to a computer because of the quantity of data and additional manual process so for simplicity we have assumed the time for data retrieval from a trail camera is double the time required for retrieval from a Cacophony Project heat camera.

The time taken to analyse the recordings captured by the cameras is calculated by averaging the analysis time across the recordings for trial period. The labour rate is assumed to be $24 per hour (including overheads).

 Chew cardsTrail camerasAI heat camera
Capital cost of devices ($)0.42502950
Useful life time of device (weeks)1300300
Number installed per hectare411
Detection rate (% compared to best)828100
Time to retrieve data0.501.000.50
Time to analyse one week of data (hours)0.0201.4200.021
Total labour per week (hours)0.5202.4200.521
Total labour cost at $24 per hour12.4858.1212.51
Weekly data storage cost ($)
Capital cost per week/hectare1.600.839.83
Total cost per week/hectare14.6265.8029.88
Cost per detection per week/hectare182.75234.9829.88

Although the AI heat camera is expensive to buy, this analysis shows that the cost per detection is dramatically lower than the other detection approaches tested. Not only is the AI heat camera the most sensitive technology for detecting possums, it is also the lowest cost.

Future trends in cost and detection

There have been significant efforts to use AI on trail camera footage. This looks promising for large mammals but appears to be less effective for identifying smaller mammals. More importantly, trail cameras miss approximately three quarters of possum sized animals. Previous trials indicate that for smaller animals, trail cameras are likely to miss an even higher proportion of interactions.

The AI heat camera is already a more sensitive, faster and lower cost way of monitoring predators and is likely to continue to improve in terms of automated identification performance and cost reduction. The camera forms part of a larger open operating system which supports rapid testing of new traps and lure designs.


Other experiments indicate that trail cameras are more sensitive than traditional monitoring tools like chew cards and tracking tunnels.This experiment shows that the Cacophony Project AI heat camera is even more sensitive again and lower cost per detection than trail cameras. It is likely that for fast moving, smaller animals the AI heat camera will be even more sensitive when compared to a trail camera.

Further research will focus on comparing various predator detection technologies in natural environments.

The experiment was funded by OSPRI and we would like to acknowledge the use of ZIP trail facilities at Lincoln