Abstract
Many people insist on getting close to wildlife without realizing how harmful the consequences of their actions can be, such as the death of the animals or even themselves. To decrease the number of such events, we aim to classify tourist photos of mountain gorillas, a critically endangered species, and other animals to detect violations derived from the distance maintained between the animals and the humans. We plan to implement object detection technology to locate the human and animal targets in the image and generate features. Then, image processing techniques are applied. We propose innovative models following multi-input logic which takes in image data and non-image features with ensemble model logic to perform the classification. We show that this model performs more accurately and has a relatively shorter run-time compared to other models when detecting violations in tourist photos with animals.
Introduction
Wildlife tourism is a double-edged sword: while it generates vital revenue for conservation efforts, it frequently leads to "distancing violations" that jeopardize both human and animal safety. Organizations like Gorilla Fund International mandate strict minimum distances to prevent disease transmission and animal stress, yet tourists often ignore these rules to capture the perfect photograph. Monitoring these interactions at scale is a significant challenge, as manually reviewing thousands of images is impractical. This paper presents a computer vision pipeline designed to automatically detect social distancing violations in tourist photos of mountain gorillas and other wildlife.
Methodology
The researchers developed a multi-stage pipeline that integrates deep learning with spatial feature analysis to move beyond simple image classification.
1. Object Detection
The first stage utilizes EfficientDet-Lite2 to identify and draw bounding boxes around humans and animals within an image. These bounding boxes provide the raw spatial coordinates necessary for downstream analysis.
2. Feature Engineering
To overcome the limitations of 2D perspective, the team engineered "non-image features" from the bounding box data. These include the distance between centroids, box areas, and area ratios, which serve as proxies for relative size and proximity.
3. Multi-Input Model Architecture
The core innovation is a dual-stream architecture:
- Image Stream: A Convolutional Neural Network (CNN) based on InceptionV3 processes raw pixel data.
- Feature Stream: A Multilayer Perceptron (MLP) processes the engineered spatial features. The outputs are concatenated and passed through a final sigmoid layer for binary classification (violation vs. non-violation).
4. Ensemble Logic
For benchmarking, a heterogeneous ensemble model was constructed, combining five distinct architectures: DenseNet 161, ResNet 50, ResNet 152, VGG 16, and VGG 19_bn.
Datasets and Stress Testing
The models were trained on a primary dataset of mountain gorilla tourist photos. To address class imbalance, synthetic non-violation examples were generated using rotation and zooming. To evaluate generalization, the models were "stress-tested" on diverse datasets, including campus statues, various internet animal photos (lions, tigers, pandas), and human-to-human social distancing scenarios.
Results and Discussion
The proposed multi-input model demonstrated exceptional performance, achieving 100% recall and 100% accuracy on the gorilla test set. The ensemble model followed closely with 98% accuracy.
Key insights from the study include:
- Spatial Context: Explicitly feeding centroid distance and area ratios provides the "depth" cues necessary for accurate distance estimation in 2D images.
- Morphological Sensitivity: While the model generalized well to animals with similar body plans (e.g., big cats), it struggled with morphologically distinct species like giraffes, indicating that "wildlife" is too broad a category for a single specialized model.
- Efficiency: The multi-input model offered the best balance of accuracy and computational efficiency, making it the most suitable candidate for real-world deployment.
Conclusion
This research demonstrates that combining deep learning image classification with engineered spatial features can effectively monitor human-wildlife interactions. Future work will focus on expanding dataset diversity, automating feature selection, and further hyperparameter optimization to enhance the robustness of the detection system across a wider array of species and environments.
Summary updated