Around 4,000 reindeer-vehicle collisions (RVCs) happen annually, leading to losses in reindeer and human lives. The annual vehicular property damage from RVCs is estimated to cost 15-20 million euros. Despite this, there have been only a limited number of attempts to automate RVC prevention. To save Rudolf and the rest of the reindeer, this paper tackles RVCs by proposing a thermal imaging-based solution that would be feasible to implement in a vehicular setting. The system relies on an inexpensive FLIR thermal camera that can be attached to most smartphones. Compared to regular cameras, thermal camera performance is not hindered by low lighting conditions or poor weather, especially winter conditions in the reindeer native lands in Lapland. For recognition of the reindeer, we utilise a Convolutional Neural Network (CNN) algorithm that we evaluate with open-sourced and self-collected FLIR thermal camera datasets. Our solution has high accuracy on diferentiating reindeer from other pedestrians, 0.998 at best, and thus compares well to other camera-based image recognition algorithms used for general animal detection. In future work, the model can be integrated into a driver warning system or an autonomous vehicle's control system to avoid the reindeer on the road and prevent collisions.
In Finland, there are around 4,000 reindeer-vehicle collisions (RVC) per year in which both reindeer
(Rangifer tarandus tarandus) and human lives are lost and property damage occurs. Annual vehicle
damage costs from RVCs are around 15-20 million euros, and reindeer owners claim circa 2.5 million
euros in insurance settlements for reindeer lost in road trafic accidents [
This work presents a thermal imaging-based method and a CNN image recognition algorithm for
detecting reindeer and diferentiating them from other pedestrians and animals, especially people
walking on the rural roadside. For the sake of automating evading movements of the vehicle, it is crucial
to know when the object is indeed a human or reindeer. We presume that a pedestrian walking the
roadside would continue progressing straight [
CEUR Workshop
ISSN1613-0073 on the roadside for prolonged periods of time. Reducing reindeer collisions is an especially pressing problem in Northern Scandinavia and Finland; they are considered culturally essential animals, an expensive delicacy (a herder owns each reindeer), and their semi-domestic nature comes without fear of trafic. In the future, we can expand to other wildlife, such as elk/moose, of which avoidance difers from tightly-packed reindeer.
Regular cameras perform well in daylight and during good weather, but bleak autumns and winters
hinder visibility and prediction accuracy in poor weather and low-light conditions. These are also the
periods when most RVCs happen. Our proposed solution combines a CNN-based image recognition
model, an of-the-shelf FLIR thermal camera, and an Android smartphone. Compared to some existing
drone-based reindeer detection methods [
Reindeer, hoofed ruminant mammals, can be classified under the family Cervidae. Diferent methods can be utilised for cervid-collision prevention. Next, we discuss reindeer-vehicle collision (RVC), or in more general terms, cervid-vehicle collision (CVC) prevention methods. Some of them include Advanced Driver Assistance Systems (ADAS) technology, but many are not integrated into the vehicles at all. Table 1 summarises some most popular CVC prevention methods, out of which we focus on digital ones. The column “Still in use?” refers to their current status in Finland.
Method Relocation Hunting Overpasses/crossings Vegetation removal (Street) lights Fencing Road signs Porokello Varoporoa Radio transmitters Night Vision Riistavaroitus
Source Still in use?
[
Yes
[
No
[
Not clear Yes
Note Not used in Finland Hunting not applicable to reindeer in Finland as they are considered livestock Located in southern parts of Finland only Done near the roads Focused near cities and settlements; thus, rural areas rarely have lights. RVCs are a higher risk in rural areas, and the road speeds are higher on rural roads.
Used in Finland for other cervids, e.g. elk/moose [
These were radio transmitters fitted on the reindeer and LED warning lights installed on the roads. They had a lot of technical issues they were working on.
Thermal camera setup connected to NV ECU. First such system in production cars.
Made for Audi, BMW, and Daimler. Used FIR or Far Infrared Camera.
Mobile application that warns the user when there is a risk for CVC. Not fullyautomatic as it requires human input.
The first example is a service explicitly provided for aiding drivers to avoid reindeer on the roads, a
smartphone application called Porokello. It was a mobile app that crowdsourced reindeer sightings.
The professional drivers of trucks and delivery vehicles were asked to log reindeer sightings in the
application, and the application would then notify other users (such as personal vehicle drivers) if there
was a heightened risk for RVCs [
A similar application to Porokello is the new Riistavaroitus application [14] that launched only recently, at the beginning of 2025. In the Play Store description1, the application is claimed to be able to warn about active reindeer areas, as well as about the end of fencing, active hunting sites, and about RVCs that have already occurred. However, such a solution still requires human input from hunters who have been notified about an RVC and who go to handle hurt animals [ 14]. Thus, there can be significant delays in issuing the RVC warnings, which reduces their information value. Another digital service is the Varoporoa website2. It is map-based, showing all reindeer vehicle crashes in the last 21 days, in addition to some historic data tables. The last year shown is 2021, with no data for 2022 or 2023. The map does not show any data at the moment. However, a separate page shows yearly RVC statistics between 2011 and 2024.
A diferent type of RVC prevention method is radio transmitter collars installed on 235 reindeer in
Saltfjellet, Norway [
Utilising thermal cameras for CVC prevention has been proposed but rarely experimented with in
real-life conditions. For example, Zhou et al. [15] built a thermal camera setup using a large FLIR
thermal camera with a frame grabber, two motors, and power sources. The work focused more on the
technical production of this thermal camera setup than on the eficiency of the algorithmic solution.
Lyo et al. [
Night Vision (NV) Forslund and Bjärkefur [
However, Riistavaroitus is the only technological solution that is still in use today. The Porokello
app was terminated, and the Varoporoa website is not available at the moment. It is unclear if NV and
the radio transmitter collar project by [
In this research, we consider animal-vehicle collisions (AVCs), more specifically cervid or reindeer vehicle collisions (RVC), that happen on public roads, including motorways, dual carriageways, and similar. The vehicles considered are all road vehicles, varying from personal cars to trucks and other heavy machinery. AVCs can happen on any road with vehicle trafic, although reindeer are more common in rural settings. Some areas also have more reindeer, and thus, the risk for RVCs is higher.
The current estimation of semi-domesticated reindeer in Finland is 203.700 animals [22]. Reindeer
are well accustomed to living in the Arctic areas with feet that have evolved to walk in deep snow. Their
movement on the roads can be determined by various factors such as the time of year or day, foraging,
or breeding season [
This research follows, on a general level, the Constructive Research Approach (CRA) [24, 25]. The focus is on creating an artefact specifically to solve a specific real-life problem, in this case, the occurrence of RVCs. As we have identified in Section 2, there is a knowledge gap on existing RVC prevention methods: 1) The current systems are either not in use anymore (most digital services) or inefective for broader applicability (such as radio transmitter collars), 2) Even if a commercial solution (Night Vision) seems to exist, there are no proofs of its performance for RVC prevention, 3) Camera-based animal-detection algorithms are problematic on nighttime when RVCs are the most common, 4) Thermal-based state of the art solutions focus on UAV perspective, leaving little applicability to real-life driving situations, and 5) No reindeer-specific image recognition algorithm has been presented in the literature that would be evaluated in dark/night environments. To fill this gap, we developed an artefact consisting of a CNN-based algorithmic solution that we validated with actual reindeer image data, combining open-sourced and self-collected images. The fit of the algorithm was evaluated with statistical metrics.
Only a limited number of thermal image open-source datasets are available, especially those that include deer images. FLIR provides its own ADAS dataset, but no deer images are included. Choi et al. [26] presented a KAIST multi-spectral dataset, but there is no information about the classes used for detection. Roboflow is an image database with 90,000 public datasets with over 220,000 images with 805 class labels [27]; three such datasets were used in this study.
Svarzas Dataset is a deer image dataset (see Figure 2, right-hand side) published by Balzs Sznyi in 2023 [28], and it is available under a CC BY 4.0 license. It contains 2,599 thermal camera images of deer in an enclosure. The images are 416 by 234 pixels. The deer are annotated using Pascal VOC format. According to the README file, some preprocessing is done to the images. The pixel data was auto-oriented, and EXIF orientation was stripped. They were also resized to 416x416 with Fit Within; image augmentation techniques were not used.
Human Detection in IR Images Dataset contains thermal camera images of humans in various poses. It was published by username Karky in 2023 [29] under the MIT license. It contains 895 images, and the humans are annotated using the COCO format. The images are 640 by 640 pixels. Just as with the first dataset by Sznyi, the pixel data was auto-oriented, EXIF-orientation stripped, and the images were resized, although here they were resized to 640x640 with Stretch. Karky did not use any image augmentation techniques, just like Sznyi.
Thermal Human Detection Dataset is published by the username PNUSafetyNet in 2022 [30]. It contains 229 thermal images of humans in diferent poses. The images are sized at 640 by 480 pixels. These images were black and white instead of the coloured other datasets. The human images were annotated using the COCO format. The only pre-processing applied was auto-orienting the pixel data and stripping EXIF orientation. Image augmentation was not used.
We collected thermal camera images of reindeer to ensure that the algorithm could generalise over more than one deer thermal image dataset, as there was a limited number of open-source datasets with deer images available and no datasets with reindeer. We used the Teledyne FLIR One Pro LT (generation 3) thermal camera attached to the OnePlus 7t Android phone. The setup is shown in Figure 3. The resolution of One Pro LT is 80 x 60 or 4800 pixels. The horizontal field of view (HFOV) is 50° ±1°, and the vertical field of view (VFOV) is 38° ±1°. There are three diferent image modes: infrared, visual, and MSX or Multi-Spectral Dynamic Imaging mode. The file formats are radiometric JPG and MPEG-4 in iOS and MP4 in Android.
We used FLIR’s own FLIR One application. The images were handheld without a tripod and in landscape and portrait orientation (see Figure 2, left-hand side). The scan was infrared only, and the colour was iron. The images were taken in Ranua Zoo. The animals were photographed in their enclosure from varying distances and angles. There were images of both single and multiple animals. The reindeer were all adults, but one juvenile was in the European Forest Reindeer enclosure. The pictures were taken during one day in October of 2023, when the weather was cloudy without rain. There were altogether 408 images, of which 135 were of the European Forest Reindeer (Rangifer fennicus fennicus), and 273 of the similar-looking domesticated reindeer (Rangifer tarandus tarandus). The size of the images was 640 by 480 pixels. The self-collected dataset is small for a machine-learning dataset, but data augmentation was used to correct this issue. We also used the open-source deer dataset from Roboflow to increase the number of deer images available for training, testing, and evaluation.
The Roboflow dataset had 842 images of deer and humans, a total of 1684 images. In the Ranua dataset, the deer images were self-collected, and the human images were collected from an open-source dataset from Roboflow. It had 281 images of deer and humans, totalling 562 images. In the mixed dataset, the deer images were a combination of images from the self-sourced and Roboflow sources, and all the human images were from Roboflow. It had 654 images of deer and humans, totalling 1308 images. The images used for model training, testing, and validation were selected from the Roboflow and Ranua datasets randomly in a 1:1 ratio. This ratio was selected to create a balanced dataset so that the model would avoid overfitting on a particular class. The image recognition models were created using Keras and KFold from Scikit-learn. 27 models were created and then compared against each other. They were created iteratively so that the evaluation metrics were looked at to see which models performed the best and how adding each layer influenced the metrics. All models except model 8 were Convolutional Neural Networks (CNNs). Model 8 was a simple Neural Network (NN) but was eventually left out of further testing as it did not perform well. Figure 4 shows model number 25.
Rescaling and data augmentation: The first layer of all 27 diferent models was a Rescaling layer, which rescaled all the images to 1. / 255. The last three layers were the Flatten layer, then two Dense layers, first with unit size 128 and activation relu, or rectified linear unit activation, and then one with unit size one and activation sigmoid. The sigmoid activation was chosen as the last one, as this was a binary classification task, and sigmoid activation provides a value between 0 and 1 for the likelihood of the image being of the chosen class. For the data augmentation, models from 9 to 27 used image augmentation layers after the Rescaling layer (see Figure 4). Image augmentation layers improve the training with small datasets [31]. The layers were RandomCrop, RandomFlip, RandomRotation, and RandomZoom, varying in diferent models. Model 9 used RandomCrop at sizes 200 and 200, RandomFlip at horizontal, RandomRotation with a factor of -0.5 and 0.5, and RandomZoom at 0.25. Model 10 used RandomRotation with a factor of -0.5 and 0.5 and RandomZoom at 0.25. Model 11 used RandomCrop at sizes 200 and 200 and RandomFlip horizontally. They both had five blocks of two Conv2D layers followed by MaxPooling2D. The rest of the models from 12 to 27 used RandomRotation with a factor of -0.5 and 0.5 and RandomZoom at 0.25.
Convolution and maxpooling: All of the convolutional layers used were square Conv2D layers, and they used padding with the same value. The activation used was relu in all of them. The number of iflters and the kernel size varied between models and sometimes between diferent convolutional layers in the same model. This was done to discover the best combination of values for identifying reindeer from thermal images. The models mostly followed the pattern of having a Conv2D layer, or several, followed by a MaxPooling2D layer. Model 1 had a Conv2D layer with 16 filters and a kernel size of 3, followed by a MaxPooling2D layer. Model 2 added another convolutional layer with 32 filters, a kernel size of 3, and a MaxPooling2D layer. Model 3 added another convolutional layer with 64 filters and a kernel size of 3 and a MaxPooling2D layer, and model 4 further added a convolutional layer with 128 iflters and a kernel size of 3 and a MaxPooling2D. The Conv2d layers in models 5-7, 9-24, and 26 had three filters and a kernel size of 3. Model 25 had the first layer with 11 filters, a kernel size of 11, five iflters, and a kernel size of 5, followed by three with three filters and a kernel size of 3 convolutional layers. Models 5-25 had five blocks of Conv2D and MaxPooling2D. In each block, models 6 and 9-11 had two Conv2D layers followed by a MaxPooling2D. The rest of the models had one Conv2D layer followed by a MaxPooling2D layer. Model 26 had seven blocks of Conv2D and MaxPooling2D combinations, and model 27 had six blocks of that layer combination.
Normalisation : In models 7, 13-20, and 25-27, normalisation layers were explored. Models 7, 13, and 25-27 used a single BatchNormalization layer after the first Conv2D layer (see Figure 4). Model 14 had the BatchNormalization layer between Conv2D and MaxPooling2d for every five blocks. Models 15 and 16 similarly used LayerNormalization, so in model 15, there was only one normalisation layer in the first convolutional block. In model 16, it was in all five blocks. Models 17 and 18 repeated the same process with UnitNormalization, and models 19 and 20 similarly used GroupNormalization. Diferent Dropout layers were explored in models 21 to 24. Model 21 used Dropout at a rate of 0.25 after each of the five Conv2D layers, and model 22 was similar, except the dropout rate was 0.5. Models 23 and 24 used SpatialDropout2D similarly; this layer came after each of the five Conv2D layers, and the dropout rate was 0.25 in model 23 and 0.5 in model 24.
After the model was created, model.compile was called with optimized adam, the loss function was set to BinaryCrossEntropy, and the chosen metrics were also given as attributes. Then model.fit was called, and the history information was taken. Test accuracy for the whole model was recorded, and precision, recall, and F1-score were recorded separately for the deer and the human classes. 10-fold cross-validation was used with varying numbers of epochs. All 27 models were initially tested on 10 epochs while training with the Roboflow and testing with the Ranua data. Out of all the models, models 13, 25, and 27 were selected for further testing. They were tested with epochs 10, 25, 50, 75, and 100. After the tests were run, the model was trained on Roboflow and tested on Ranua data. Models 13, 25, and 27 were run with the same epoch numbers while training with Ranua and testing with Roboflow data, as well as with a mixed dataset. 25% of the data was reserved for testing.
The test accuracy results for all models with 10 epochs are shown in Figure 5. The results vary widely
between 0.447 and 0.998. When considering that 0.5 means that the accuracy is 50%, which is the
accuracy of a guess, it can be seen that some of these models are as good as a guess. However, there are
several models with very good accuracy results. Out of the 27 models, as mentioned in the previous
section, models 13, 25, and 27 were chosen for further tests. Some model-epoch combinations might
have poor test accuracy, poor recall, and F1-score, but precision is 1, or poor test accuracy, precision,
and F1-score, but recall is 1. Thus, it is necessary to look at all the metrics. Our best accuracy results
are close to 0.998, which aligns with animal recognition algorithms presented in the state of the art,
even if they use an ”easier” case of the camera instead of thermal images [
Table 2 and Figure 6a show the test accuracy for training with the Roboflow and testing with the Ranua data. The test accuracy varies between 0.493 (model 27 with 100 epochs) and 0.998 (model 13 with 10 epochs). However, 10 of the 15 diferent models and epoch combinations had an accuracy above 0.85. Table 2 also has the precision, recall, and F1-score. Most models with low test accuracy also had low precision, recall, and F1-scores. The models with better accuracy perform better and best with ten epochs. For example, model 13 at 10 epochs had the best test metrics out of model and epoch combinations with an accuracy of 0.998, with deer precision of 1, recall of 0.996, and F1-score of 0.998, and the values for the human class of 0.996, 1, and 0.998, respectively.
Increasing the epoch count seems to be somewhat inconsistent with the metrics. The test accuracy drops when the epochs increase from 10 to 25 for all three models. They further drop for models 13 and 25 when the epochs are increased from 25 to 50, although for model 27, the accuracy improves. Accuracy seems to improve or stay the same when the epochs are further increased to 75. Yet, they drop again for models 25 and 27 when the epoch count is increased to 100, but model 13 performed well. It had an accuracy of 0.979, which is better than model 25 at epoch 10. Our results compare well with previous work on animal recognition. Kalla et al. [32] presented Nasnet, a CNN-based Nasnet-Mobile model, for identifying horses on the roads. Their training accuracy was approaching 1 at 6 epochs, and for testing accuracy, it was approaching 1 at 45 epochs. Their precision was around 86.5%. When comparing these results to our results shown in Table 2, ours is higher for all the models with 10 epochs when the range is between 0.946 and 1, but there is more variation with higher epoch counts.
Antônio et al. [33] presented K-Nearest Neighbors (KNN) and Random Forest (RF) algorithms for animal detection. They reported their results using F-measure. They had two results for KNN (0.611 and 0.624) and two for RF (0.560 and 0.589). 25 out of 30 of the F1-scores shown in Table 2 are higher than the results from Antônio et al. 20 of the results have an F1-score above 0.8. Nguyen et al. [34] presented a CNN for detecting and identifying wild animals from images. They used a simplified version of AlexNet called Lite AlexNet, VGG-16, and ResNet-50. They tried both imbalanced and balanced datasets. For the binary task of animal recognition, the accuracy results were for VGG-16 96,6% for imbalanced and 95.9% for balanced, for ResNet-50 they were 96.1% and 95.7% respectively, and for Lite AlexNet they were 94.9% and 92.7% respectively. For wildlife identification tasks, accuracy varied between 82.5% for Lite AlexNet for identifying the six most common species and 90.4% for ResNet-50 for identifying the three most common species with an imbalanced dataset [34]. Table 2 shows that all three of our best models with epoch 10 have better accuracy.
The accuracy metrics for the model epoch combinations when training with Ranua and testing with Roboflow data are shown in Figure 6b and Table 3. The test accuracy varied between 0.5, which several model epoch combinations had, and 0.88, with model 27, with 10 epochs. From Table 3 we can see that the best accuracy is for model 27 with 10 epochs at 0.88, and the following best are for model 25 with 25 epochs at 0.859 and model 25 with 100 epochs at 0.855. The same happens in test accuracy, and performance correlates with other metrics. However, several model epoch combinations might have perfect precision or recall for one class but a much lower value for the other.
Compared our results to previous works, Kalla et al. [32] have better accuracy than any model or epoch combination in Table 3. When comparing the precision of 86.5% from Kalla et al. [32], we have better precision in model 27 with 10 epochs for deer (0.9) and almost equal for humans (0.862). Despite the lower accuracy, the precision can compare well for model 27 at 10 epochs. When comparing our F1-scores to the F-measure reported by Antônio et al. [33] (ranging between 0.5601 and 0.6243), from our results, 20 out of 30 F1-scores are higher. The F1 scores, however, are not as high for the model here as for the one trained with Roboflow and tested with Ranua. The accuracy values for Nguyen et al. [34] with a range of 92.7% to 96.6% for the binary identification were all better than in Table 3.
The results for the mixed dataset are very good overall, as shown in Table 4 and Figure 6c. The test accuracy varies between 0.905 (model 25 with 50 epochs) and 1, with 11 of the 15 model epoch combinations having this. As the combined dataset is quite large, it is probably one cause for the better performance compared to the results from training with Ranua and testing with Roboflow. The results compare well with related literature. The test accuracy is 1 for 11 out of the 15 model and epoch combinations; at worst, it is 0.905. It compares well to the test and training accuracy 1 by Kalla et al. [32]. The precision results are superior compared to the 86.5% by Kalla et al. The F1-scores are all higher by a large margin than the F-measure results for Antônio et al. [33]. When compared to Nguyen et al. [34], it can be seen that all but one of our accuracy results in Table 4 are higher.
As can be seen from the results, the mixed dataset performed the best out of the three, and the models that were trained on Roboflow and tested on Ranua material were the 2nd best. One possible reason for the lower results in the case of models trained on Ranua and tested on the Roboflow data could be the small size of the training material. On the other hand, size is not necessarily the only thing, as the mixed dataset performed better than the one trained with Roboflow, tested with Ranua, despite the latter being a larger dataset (2 times 842 images vs. 2 times 654). As the results varied between the models trained on diferent datasets, it is hard to recommend the best epoch counts. However, it seems that high epoch counts are not needed to attain well-performing models. The limitation of this work is that it only has two classes. This makes it easier to identify reindeer from images, as only reindeer and human images are included. However, our future work includes adding more classes to the model. At least elk or moose would be an essential addition, but other animal classes could include fox, rabbit, cat, and dog. If the detection of diferent Arctic wildlife, including birds, were enabled, then the data collection could serve a secondary purpose of collecting critical information about the occurrence of rare species. This could be used to help conservation eforts. However, including additional classes creates a need to gather a larger dataset, as thermal images of varying species are sparse in the open-source datasets.
Another future work is to gather more thermal camera images of reindeer and humans. These could also be taken in a driving situation, providing the most realistic images. Increasing the dataset size would improve the model, as machine learning applications usually perform better with larger datasets. Increasing the dataset with additional reindeer and human pictures, as well as pictures for additional classes, would then likely afect the model training time. However, model complexity and epoch counts could potentially be optimised for better time performance, as shown in our work. Our best evaluation metrics for the models were so good with epoch 10 and reasonable processing times. As such, our future work will focus more on the datasets and real-time animal detection situations. In the future, our goal is to also do realistic testing of the model in a vehicular setting. This would provide the most realistic testing results, but would require a test setup where the thermal camera can be safely positioned outside of the test vehicle. An additional challenge is where to find reindeer for testing purposes while driving, as they tend to appear only at times when you would wish they did not.
In this paper, we presented a CNN model that utilised thermal images to identify reindeer and diferentiate them from other pedestrians, specifically people. The aim was to create a system that, in the future, could be integrated into a driver warning system or into an autonomous vehicle’s control systems to avoid reindeer-vehicle collisions. The model was trained on an open-source dataset and tested on the self-sourced one, and vice versa, as well as trained and tested on a mixed dataset. The evaluation shows good comparison against and, at times, exceeds the evaluation metrics of previous research. The best accuracy of 0.998 was a promising result for reindeer identification from thermal images. Most importantly, we can showcase that thermal imaging is equally accurate to standard camera-based methods when it comes to the identification of reindeer. However, typical cameras do not perform well on dark nights, and our solution with thermal imaging overcomes this restriction.
The work has been supported by the EU HORIZON projects CHIPS-JU CIA FEDERATE (grant 101139749) and CHIPS-JU RIA HAL4SDV (grant 101139789), Business Finland HAL4SDV national funding (grant 7655/31/2023), and the Finnish Research Council project Northern Utility Vehicle Laboratory Consortium GO!-RI (grant 352726). The authors are thankful to Ranua Resort.
The authors have not employed any Generative AI tools. [13] I. Hansen, S. M. Eilertsen, G. H. M. Jørgensen, J. Karlsson, The “animal sense” warning system.
low-cost technology to prevent collisions between semi-domestic reindeer and vehicles, 2021. [14] Yle, Uusi mobiilisovellus varoittaa hirvistä reaaliaikaisesti – hälyttää myös käynnissä olevasta metsästyksestä, Yle Uutiset (2025). URL: https://yle.fi/a/74-20130962, accessed: 2025-04-04. [15] D. Zhou, M. Dillon, E. Kwon, Tracking-based deer vehicle collision detection using thermal imaging, in: Int Conf on Robotics and Biomimetics, 2009, pp. 688–693. doi:10.1109/ROBIO.2009.5420589. [16] H. Lyu, F. Qiu, L. An, D. Stow, R. Lewision, E. Bohnett, Advancing wild deer monitoring through uav thermal imaging and modified faster rcnn: A case study in nepal’s chitwan national park, Sensing and Imaging 25 (2024) 50. [17] J. Friedman, T. Hastie, R. Tibshirani, Additive logistic regression: a statistical view of boosting,
The annals of statistics 28 (2000) 337–407. [18] R. E. Schapire, The boosting approach to machine learning: An overview, Nonlinear estimation and classification (2003) 149–171. [19] P. Viola, M. Jones, Rapid object detection using a boosted cascade of simple features, in: Computer
Society Confon Computer Vision and Pattern Recognition, volume 1, IEEE, 2001. [20] L. Masello, G. Castignani, B. Sheehan, F. Murphy, K. McDonnell, On the road safety benefits of advanced driver assistance systems in diferent driving contexts, Transportation research interdisciplinary perspectives 15 (2022) 100670. [21] B. Sznyi, szarvas dataset, https://universe.roboflow.com/balzs-sznyi-vqd7w/szarvas, 2023. URL: https://universe.roboflow.com/balzs-sznyi-vqd7w/szarvas, visited on 2024-08-28. [22] Paliskunnat, Poromäärien kehitys, 2024. URL: https://paliskunnat.fi/py/materiaalit/tilastot/ poromaarien-kehitys/. [23] Paliskunnat, Poro (rangifer tarandus tarandus), 2024. URL: https://paliskunnat.fi/poro/poro/. [24] K. A. Piirainen, R. A. Gonzalez, Seeking constructive synergy: Design science and the constructive research approach, in: Design Science at the Intersection of Physical and Virtual Design: 8th International Conference, DESRIST 2013, Helsinki, Finland, June 11-12, 2013. Proceedings 8, Springer Berlin Heidelberg, 2013, pp. 59–72. [25] K. Lukka, Konstruktiivinen tutkimusote: luonne, prosessi ja arviointi, in: K. Rolin, M.-L. KakkuriKnuuttila, E. Henttonen (Eds.), Soveltava Yhteiskuntatiede ja Filosofia, Gaudeamus, Helsinki, Finland, 2006, pp. 111–133. [26] Y. Choi, N. Kim, S. Hwang, K. Park, J. S. Yoon, K. An, I. S. Kweon, KAIST multi-spectral day/night data set for autonomous and assisted driving, IEEE Trans on Intelligent Transportation Systems 19 (2018) 934–948. doi:10.1109/TITS.2018.2791533. [27] F. Ciaglia, F. S. Zuppichini, P. Guerrie, M. McQuade, J. Solawetz, Roboflow 100: A rich, multidomain object detection benchmark, arXiv preprint arXiv:2211.13523 (2022). [28] B. Sznyi, ”szarvas dataset”, 2023. URL: https://universe.roboflow.com/balzs-sznyi-vqd7w/szarvas. [29] Karky, Human detection in ir images dataset, 2023. URL: https://universe.roboflow.com/ karky-bk9o4/human-detection-in-ir-images. [30] PNUSafetyNet, Thermal human detection dataset, 2022. URL: https://universe.roboflow.com/ pnusafetynet/thermal-human-detection-00eyg. [31] C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of big data 6 (2019) 1–48. [32] H. Kalla, B. Ruthramurthy, S. Mishra, G. Dengia, R. Sarankumar, A practical animal detection and collision avoidance system using deep learning model, in: Int conf for Convergence in Technology (I2CT), IEEE, 2022, pp. 1–6. [33] W. H. Antônio, M. Da Silva, R. S. Miani, J. R. Souza, A proposal of an animal detection system using machine learning, Applied Artificial Intelligence 33 (2019) 1093–1106. [34] H. Nguyen, S. J. Maclagan, T. D. Nguyen, T. Nguyen, P. Flemons, K. Andrews, E. G. Ritchie, D. Phung, Animal recognition and identification with deep convolutional neural networks for automated wildlife monitoring, in: Int Conf on Data Science and Advanced Analytics (DSAA), IEEE, 2017, pp. 40–49.