In this study, an unmanned aerial vehicle (UAV) captures images of wild animals at different altitudes in order to compare the individual and herd identification capabilities of custom YOLO v3 and v4 models. Previous studies showed that UAVs can disturb animals in the wild at certain altitude, such that it is necessary to maintain altitude that does not disturb them. However, as the UAV altitude increases, the captured images lose features critical to YOLO in classification. We investigate and compare the accuracy of custom YOLO v3 and v4, especially from the acceptable minimum altitude and higher. We studied eight classes of wild African animals, namely, individual and herd of giraffes (Giraffa camelopardalis), individual and herd of white rhinos (Ceratotherium simum), individual and herd of wildebeests (Connochaetes taurinus), and individual and herd of zebras (Equus quagga). As UAV altitude increased, some image features are lost resulting to a model detection accuracy as low as 68.86%. The customised YOLO v4 model has 51.70 FPS outperforming customised YOLO v3 by an increased model speed of 13.7%.

This paper presents a method to classify human activities as normal or suspicious using YOLOv3 to automatically process video footages taken from a high altitude moving aerial camera, such as the one attached to a drone. We consider four human activities namely, jogging, walking, fighting, and chasing. Objects generally appear much smaller, with less visible features, when viewed from high altitudes. The reduced visible features make automatic human activity detection from ground surveillance cameras not applicable to the high altitude case. Through transfer learning, we modified existing pre-trained YOLOv3 convolutional neural networks (CNN`s) and retrained with our own high aerial human action dataset. By so doing, we were able to customize YOLOv3 to detect, localize, and recognize aerial human activities in real-time as normal or suspicious. The proposed approach achieves a promising average precision accuracy of 82.30%, and average F1 score of 88.10% on classifying high aerial human activities. We demonstrated that YOLOv3 is a powerful approach and relatively fast for the recognition and localization of human subjects as seen from above.

Background/Objectives: This study investigated the use of artificial intelligence (AI) to identify and match barefoot prints belonging to the same individual on soft and sandy soil substrates. Recognizing footprints on soil is challenging due to low contrast and variability in impressions. Methods: We introduce Deep Learning Footprint Identification Technology (DeepFIT), based on a modified You Only Look Once (YOLOv11s) algorithm, using three methods, namely, Bounding Box (BBox), 16 anatomical landmarks, and automatically segmented outlines (Auto-Seg). An Extra Small Detection Head (XSDH) was added to improve feature extraction at smaller scales and enhance generalization through multi-scale supervision, reducing overfitting to specific spatial patterns. Results: Forty adults (20 males, 20 females) participated, with 600 images per individual. As the number of individuals in model training increased, the BBox model’s accuracy declined, resulting in misclassification on the test set. The average performance accuracy across both substrates was 77% for BBox, 90% for segmented outlines, and 96% for anatomical landmarks. Conclusions: The landmark method was the most reliable for identifying and matching barefoot prints on both soft and sandy soils. This approach can assist forensic practitioners in linking suspects to crime scenes and reconstructing events from footprint evidence, providing a valuable tool for forensic investigations.

This paper investigates human–AI collaboration through explainable AI where we interpret the results of barefoot print clustering using unsupervised machine learning. This can be used to identify the number of individuals from barefoot prints on the ground as a tool in forensics or anti-poaching. A self-supervised vision transformer, DINOv2, is used to automatically extract feature embeddings from localized barefoot-print regions to identify trails belonging to an individual on soil substrate. Furthermore, we introduce an Embedding Spatial Attribution Module (ESAM) to generate spatial attribution heatmaps, enabling visualization of discriminative regions that contribute to individual-specific trail identification and improving model explainability. The proposed method is named XTrail-ID, an explainable human footprint trail identification framework with two variants, OBB-XTrail-ID (oriented bounding box-based), and SEG-XTrail-ID (segmentation-based). We quantify embedding similarity using three complementary metrics: cosine similarity, Pearson correlation coefficient, and Spearman rank correlation. Twenty adults (ten males, ten females) participated, with a total of 1000 trail images extracted from UAV imagery. SEG-XTrail-ID using cosine similarity yielded the highest performance, with (3.21) discriminability and (94.2%) accuracy, while OBB-XTrail-ID using cosine similarity achieved (2.54) discriminability and (91.5%) accuracy. In addition, the latter exhibited reduced consistency in footprint grouping when more than three individuals were present within a single frame.