A step change towards ‘precision’ wildlife control through automatic wild boar age classification - 1st Workshop on Wildlife Fertility Control: What Now? What Next? Where To? June 17 – 18, 2024- University of York - York, United Kingdom

Author(s)
Riccardo Primi
Date Issued
2024
Type
conferenceObject
Abstract
Distributing contraceptives in baits to regulate wildlife fertility or activating traps for capturing animals pose challenges, including selectivity issues. Feasibility and effectiveness depend on precise species identification, as well as discerning sex and age classes. Although some species-specific devices are available, they are mechanically triggered by the animal, lacking certainty in absolute selectivity. We employed the Ultralytics’ YOLOv8 object detection algorithm to automate wild boar detection, including age classification, which is also valuable for automating bait dispenser and traps selectivity. The model was trained, validated, and tested using a dataset consisting of 4865 images captured by camera traps placed across Central Italy. These images were annotated for species and age classes (piglets, juveniles, adults) by three experienced operators using the Roboflow platform. The class retained for the analysis was the one identified by at least two out of the three labellers. The training-validation-test ratio was 7:2:1, with data augmentation used to address diversity and mitigate overfitting. With 9876 images utilized, YOLOv8 ran on a laptop with an NVIDIA® GPU RTX™ 3070 Ti with 8 Gb vRAM. The probability of correct age classification was 0.76, with misclassifications primarily affecting subjects at the class boundary. In such cases, direct observation by skilled personnel could also yield dubious results. Testing on a NVIDIA® Jetson Nano™ board for real-time applications, such as wild boar identification for fully automating the triggers of selective traps and contraceptive dispensers, yielded successful results. In fact, power consumption of 2.3 Watt and detection time of 5 frames/s are indeed fully compatible with in-field, precision real time applications
Handle
Conference(s)
1st Workshop on Wildlife Fertility Control: What Now? What Next? Where To? June 17 – 18, 2024 University of York - York, United Kingdom