This project investigates the effectiveness of YOLOv5, YOLOv8, and YOLOv11 models in detecting persons in Wilderness Search and Rescue (WiSAR) scenarios using both RGB and thermal infrared (IR) imagery. The aim is to evaluate model performance across detection accuracy, inference speed, and generalization capability in natural, occluded environments.
WiSAR missions face obstacles such as dense vegetation, lighting variability, and environmental occlusion, which reduce the effectiveness of manual search efforts. Deep learning methods like YOLO have shown potential for automating person detection. Leveraging both RGB and IR data can enhance detection under varied conditions (e.g., daylight and low visibility). This project focuses on the WiSARD dataset, which contains labeled RGB and thermal image pairs specific to SAR contexts.
We evaluated YOLOv5, YOLOv8, and YOLOv11 for human detection in wilderness search and rescue (WiSAR) using the WiSARD dataset. Each model was trained and tested separately on RGB and thermal IR image subsets, selected from the “MtErie” and “FHL” locations. A 60-20-20 split was applied within each folder, and default YOLO data augmentation was enabled. Training was conducted in Google Colab with manual hyperparameter tuning (batch size, learning rate), using Ultralytics' implementation and custom YAML configuration for single-class detection.
- YOLOv5: Baseline model widely used in SAR literature and the WiSARD paper.
- YOLOv8: Balanced in terms of speed, accuracy, and hardware efficiency.
- YOLOv11: Recent architecture shown to handle occlusion and difficult detection scenarios effectively.
- Older or less-documented versions (e.g., YOLOv6, v7, v9, v10) were excluded due to inconsistent results or poor SAR relevance.
We used the WiSARD dataset, a multimodal collection of over 70,000 RGB and thermal infrared images captured via UAVs in wilderness environments. For this project, we selected images from the MtErie and FHL locations based on criteria including human presence, occlusion, and alignment of RGB–IR pairs. Only labeled images were included. The dataset was split 60-20-20 (train/val/test) per folder to ensure environment diversity and mitigate aggregation bias.
Refer to the link below for more information about the WiSARD Dataset: https://sites.google.com/uw.edu/wisard/
- Data preprocessing: Image resizing, normalisation, and formatting into YOLO-compliant annotations.
- Thermal-visual separation: Models were trained separately on RGB and IR subsets to compare modality performance.
- Training & tuning: Uniform hyperparameters were applied initially; adjustments were made based on results and resource constraints.
- Evaluation: Recall, precision, and mAP were used as performance metrics.
- Ethical Considerations
We used the FATES framework to evaluate ethical issues around dataset bias, human subject visibility, and transparency in model reporting.
Addressing our research question for determining the best recall performance in both modalities and model versions, YOLO v11 using IR images outperformed the other models in recall (0.592), meaning the model could identify 59.2% of all True Positives (TP) in the dataset. However, as shown in Figure above, this was only a slight improvement of 0.013 in comparison to YOLO v8 with IR which achieved a recall score of 0.579.
- Thermal IR consistently outperformed RGB across all metrics and YOLO versions.
- YOLOv11 with thermal IR achieved the highest recall (0.592).
- YOLOv11 was the top performer in both RGB and IR modalities.
- Manual hyperparameter tuning improved stability across models.
- Aggregation bias was identified and mitigated by ensuring folder-level distribution in train/val/test sets.
All training logs, configuration files, and visualisations are included in the /results and /configs directories.
Our review of existing studies found Gupta and Singh (2025) to be the most closely aligned with our work. It’s the only study with similar aims and objectives, directly comparing YOLO versions for person detection while reporting recall values. However, it differs in that they only investigated YOLO model versions one through seven and trained their models using RGB data. Despite this, it remains the most relevant study as we can directly compare our findings to theirs.
Gupta and Singh (2025) reported that version seven gave their greatest recall score. This result compares to our findings for research question two (RQ2), which found version eleven to be the best YOLO model version for RGB data training. Therefore, our results are similar, as both found the most recent YOLO model version to perform best in terms of recall score when trained on RGB data. Our study integrates thermal IR data into model training, reflecting the growing recognition of the limitations of relying solely on RGB images. For instance, Bañuls et al. (2020) showed that combining RGB and thermal IR images improves detection under challenging conditions. While we trained and tested models separately on each modality, our results indicate that IR improves recall rates, especially when paired with more advanced YOLO versions. This supports the idea that newer YOLO architectures and thermal data training enhance AI-based human detection in wilderness search and rescue scenarios. Further research is needed to explore multimodal fusion of RGB and thermal IR images to optimise accuracy for these missions.
- Ikmal Basirun (Project Manager): Led planning, coordination, team meetings, dataset aggregation and review, model development.
- Amira Faisal: Led background research and literature review, contributed to dataset review, designed project poster.
- Zahra Husain: Contributed to literature review and dataset evaluation.
- Thomas Reeve: Supported literature review, data aggregation, and model development.
- Faisal Alharbi: Focused on hyperparameter tuning research, assisted in model development.
- Syahmina Mohd Ansar: Contributed to ethical considerations (FATES) and supported documentation.
- Gupta, P., & Singh, U. (2025). Evaluation of several yolo architecture versions for person detection and counting. Multimedia Tools and Applications. https://doi.org/10.1007/s11042-025-20662-z
- Bañuls, A., Mandow, A., Vázquez-Martín, R., Morales, J., & García-Cerezo, A. (2020, November 1). Object Detection from Thermal Infrared and Visible Light Cameras in Search and Rescue Scenes. IEEE Xplore. https://doi.org/10.1109/SSRR50563.2020.9292593