This thesis addresses the recognition of free-ranging wildlife in trail-camera im-
ages using transfer learning and the YOLOv11s model. The objective was to
develop a model capable of detecting an animal, locating it with a bounding box,
and assigning it to the appropriate class. The study also examined how con-
verting daytime colour images to grayscale affects detection performance and the
distinction between female and male roe deer.
A custom dataset containing daytime RGB and nighttime infrared images was
collected with trail cameras in the Semiˇc area. The images were annotated using
the Roboflow platform, and data augmentation increased the dataset to 3014
images. Two models based on a pretrained YOLOv11s network were trained in
the Kaggle environment. The first model used a combination of daytime RGB and
nighttime IR images, while all images used by the second model were converted
to grayscale. Both models used the same classes and data split.
On the validation set, the RGB and IR model achieved a precision of 0.771,
a recall of 0.672, an mAP@0.5 of 0.803, and an mAP@0.5:0.95 of 0.644. The
grayscale model achieved a precision and recall of 0.756, an mAP@0.5 of 0.777,
and an mAP@0.5:0.95 of 0.621. The grayscale model achieved a higher recall and
a higher maximum F1 score, whereas the RGB and IR model achieved higher
mAP values and better results when distinguishing between female and male roe
deer. The results indicate that removing colour information did not improve over-
all performance. For more reliable practical use, the dataset should be expanded,
balanced, and evaluated on images from additional locations.
|
