The motivation is stark. Wild mammal populations have declined by more than 70 percent between 1970 and 2020 according to WWF figures cited in the study, and wild mammals now account for only about 6 percent of total mammal biomass, with humans and livestock making up the rest. At the same time, collisions between vehicles and animals are rising across Europe as ungulate populations and traffic density grow. In 2025 alone, more than 14,500 moose and deer were killed on Norwegian roads and railways. In the United States, an estimated 725,000 to 1.5 million animal-vehicle collisions are reported to police each year, while Canada records 4 to 8 animal deaths per day. Automated detection from roadside cameras could one day feed alert systems that warn drivers before an animal steps onto the road.

The core of Hjort-YOLO is a targeted re-engineering of YOLOv11n, the nano variant of the popular You Only Look Once family of single-stage detectors. Single-stage models predict object position and class in one pass, which makes them far faster than two-stage detectors such as Faster R-CNN, where regions of interest are proposed first and classified second. YOLOv11n was chosen as the baseline precisely because its low computational complexity suits real-time applications on resource-constrained edge devices. The team then modified the architecture in three coordinated ways: they replaced the C3k2 blocks in the lower backbone and neck with a refined cross-stage partial convolutional block called C2f-W, they swapped the Spatial Pyramid Pooling-Fast module for an area-attention block named A2C2f-W, and they removed one of the three detection heads.

Each modification addresses a specific weakness. The C2f-W block, inspired by the C2f design in YOLOv8, splits input features into two streams, passes one through bottleneck blocks, and concatenates the outputs, creating dense gradient paths that improve channel interaction and gradient propagation for higher-level representations. The A2C2f-W block, adapted from the YOLOv12 architecture, computes attention not pixel by pixel but over predefined areas of the feature map. This area-attention strategy captures mid-range contextual relationships while avoiding the heavy computational cost of full self-attention, which is crucial for identifying complex and partially occluded animals in cluttered forest scenes. The removal of the 80 by 80 pixel detection head, which normally handles small objects, reflects a deliberate design decision: camera traps and roadside sensors capture animals at relatively close range, so targets are large in the frame and the smallest-object head adds cost without proportional benefit.

The framework was evaluated on two datasets. The first is a custom Svanøy deer dataset of 3,183 images, comprising 1,432 daytime RGB images and 1,751 near-infrared monochrome images captured at night by Browning trail cameras with infrared flash, all labeled manually in Roboflow. The second is the public African Wildlife dataset of 1,504 RGB images covering four classes: buffalo, elephant, rhino, and zebra. Both datasets were split 80-10-10 into training, validation, and test sets, and all models were trained from scratch without pretrained weights under identical conditions on a Google Colab Tesla T4 GPU, using stochastic gradient descent with 300 epochs, early stopping patience of 20, and 640 by 640 pixel inputs.

The results show a favorable trade-off between accuracy and speed. On the Svanøy validation set, Hjort-YOLO achieved a mean average precision of 0.981, a precision of 0.975, a recall of 0.950, and an F1-score of 0.962, edging out the YOLOv11n baseline at 0.979 mAP. More striking are the efficiency gains: inference time dropped from 8.0 to 6.8 milliseconds per image, throughput rose from 21.98 to 24.98 frames per second, and computational complexity measured in GFLOPS fell 19 percent, from 6.4 to 5.1. The model has 2.90 million trainable parameters and a 6 MB weight file, only marginally larger than YOLOv11n’s 2.59 million parameters and 5.5 MB file, while the network was slimmed from 181 to 143 layers. Ablation studies confirmed that each component contributes: removing the third detection head alone cut GFLOPS from 6.4 to 4.6, and adding the C2f-W and A2C2f-W blocks restored accuracy while keeping complexity low. A variant retaining all three heads reached the highest mAP of 0.990 but required 6.8 GFLOPS, so the two-head design sacrifices less than one percentage point of accuracy for a 25 percent reduction in computation.

Training across ten repeated runs demonstrated robustness, with a mean mAP of 0.9804 plus or minus 0.004 and consistently small standard deviations across precision, recall, and F1-score. Notably, the combined RGB and infrared dataset outperformed models trained on either image type alone, with mAP of 0.979 compared to 0.969 for RGB-only and 0.961 for infrared-only, suggesting that visual diversity across illumination conditions improves generalization. On the African Wildlife dataset, Hjort-YOLO outperformed both YOLOv8n and YOLOv11n among lightweight models with an overall mAP of 0.922, achieving the best rhino mAP50-95 of 0.804, though zebra remained the hardest class, likely due to fewer training samples, smaller object size, and occlusion. Compared with the transformer-based RF-DETR Nano, which matched the mAP at 0.981 but required 31.5 GFLOPS, roughly five times the baseline, Hjort-YOLO stands out for delivering comparable accuracy at a fraction of the computational cost.

What distinguishes Hjort-YOLO from many lightweight detectors is its emphasis on explainability. The researchers integrated two explainable artificial intelligence tools as post-processing steps. Grad-CAM generates heatmaps from the final convolutional layer, highlighting which regions of the image most influence the prediction. Captum, a PyTorch-compatible interpretability library, applies attribution methods such as Integrated Gradients, GradientShap, DeepLift, and Occlusion to trace which pixels within the predicted bounding box drove the objectness score. Because the removed high-resolution detection head had contributed redundant gradients and background noise, the Captum heatmaps proved cleaner and more precise than Grad-CAM’s whole-image maps, revealing meaningful animal structures and helping identify irrelevant background features that could trigger false positives. This transparency matters for conservation applications, where understanding why a model fires can expose bias and guide architecture improvements.

The framework also pairs detection with segmentation. Hjort-YOLO’s bounding boxes serve as geometric prompts for the Segment Anything Model 2, or SAM2, which refines each detection into a pixel-accurate instance mask through a vision transformer and cross-attention decoder. This two-stage detection-segmentation pipeline isolates animals from background vegetation, reducing ambiguity and improving visualization. On the T4 GPU, the full end-to-end pipeline of detection, segmentation, and explainability took 607.61 milliseconds per image, with Hjort-YOLO itself consuming only 10.13 milliseconds of wall time, while SAM2 required 267.10 milliseconds and the XAI stage 321.69 milliseconds. The authors suggest that the heavy visualization stages could run on a connected edge device or cloud server, while the fast detector operates locally.

The researchers are candid about limitations. The model is tuned for large targets and may struggle with small animals; the training data comes from forested Svanøy Island rather than actual roadside scenes, so real-world deployment near roads remains untested; and benchmarking on physical embedded hardware such as Jetson Nano or Raspberry Pi is still needed to confirm real-time performance, memory footprint, and deployment overhead. Future work includes implementing the model on embedded devices as part of a driver alert system, exploring transformer-based architectures such as RF-DETR for further efficiency, and adding multi-frame object tracking combined with other detection modalities to improve reliability. For now, Hjort-YOLO demonstrates that carefully targeted architectural surgery, rather than sheer model size, can deliver real-time wildlife detection that is both accurate and interpretable on the kinds of low-power hardware where conservation technology actually lives.

Subject of Research: A lightweight, explainable YOLO-based computer vision framework for detecting wildlife in camera-trap images

Article Title: Hjort-YOLO: a lightweight and explainable computer vision framework for wildlife detection

Article References: Momin, M. A., Khan, M., Arghandeh, R., Bischof, R., Hegland, S. J., Moen, T., & Stølen, M. F. (2026). Hjort-YOLO: a lightweight and explainable computer vision framework for wildlife detection. Frontiers in Artificial Intelligence, 9, Article 1930016. https://doi.org/10.3389/frai.2026.1930016

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1930016

Keywords: Hjort-YOLO, computer vision, wildlife detection, YOLOv11n, camera traps, deep learning, explainable AI, animal-vehicle collisions, object detection, edge computing, SAM2 segmentation, Norway

The study, led by Yanchao Wang and Kaiying Wang with colleagues at institutions in Zhejiang Province, was published in the journal Smart Agricultural Technology. The team built their system around a modified deep learning architecture they call HSM-YOLO, short for Hu Sheep Mounting-YOLO, named after the Hu sheep breed they studied, which is recognized in China as a national-level protected livestock genetic resource. Their approach addresses a persistent gap in precision livestock farming: while computer vision systems for detecting mounting behavior in cattle and pigs have advanced rapidly, sheep have received comparatively little attention, and existing sheep models tend to be bulky, computationally expensive, and tested only in single, fixed farming environments.

The challenge the researchers set out to solve is harder than it might appear. In dense group-housed environments, sheep exhibit complex social interactions that include not only sexual mounting related to estrus but also non-sexual mounting triggered by food competition, social play, and stress responses. Both behavioral patterns look strikingly similar on camera: one animal places its forelimbs on the back of another, stands on its hind limbs for support, and maintains a forward-leaning posture. The crucial distinction lies in the details. Non-sexual mounting typically lacks the rhythmic pelvic thrusting movements of reproductive mounting and shows arbitrary forelimb positioning, manifesting primarily as simple dorsal pressure. These subtle postural differences are exactly what the new model was designed to learn.

To train and test the system, the researchers collected video from two commercial sheep farms in Huzhou City between April 2024 and February 2025, using healthy Hu sheep aged six to twelve months. Rather than relying on a single camera setup, they deliberately engineered four distinct imaging scenarios to push the model’s adaptability. Five dome cameras installed three meters above ground captured footage of side pens near windows, where fluctuating light, fence bars, and sheep clustering created interference. Nine cameras at 2.75 meters monitored the central pens, which suffered from insufficient illumination, close-range occlusions, and animals temporarily exiting the field of view. Three cameras of a different model provided supplementary imaging, and an inspection robot circling the entire barn at 0.3 meters per second added motion blur to the mix of challenges.

From this footage, the team manually selected 867 video clips ranging from roughly three to twenty seconds, then extracted frames and applied a perceptual hashing algorithm to eliminate redundant near-identical images. The result was a curated dataset of 6,655 images divided into training, validation, and test sets at an approximate six-to-two-to-two ratio. Crucially, the annotation scheme was innovative: rather than treating all mounting as a single category, veterinary-supervised labeling divided the interactions into three classes: sexual mounting, non-sexual mounting, and the mounted ewe herself. Non-sexual mounting was introduced primarily to absorb interference, so labels for mounted individuals were not required for that category. This three-class structure allows the model to simultaneously recognize the behavior and localize the fertile animal.

The technical heart of HSM-YOLO lies in four coordinated modifications to the YOLOv11n architecture, one of the most computationally efficient models in the YOLO family. First, the researchers replaced most standard convolutional layers in the backbone network with depthwise separable convolutions, a design that splits feature extraction into a depthwise operation applying a single filter per input channel and a pointwise operation working across channels. The mathematics is compelling: where a standard convolution requires k²×n×m parameters, the separable version needs only k²×n plus n×m, a compression ratio of one over m plus one over k². This dramatically cuts parameters and computation while preserving feature representation, reserving computational overhead for the more demanding modules that follow.

Second, the team upgraded the C3k2 blocks in the backbone to a novel dual-path module called C3k2_BiRep, which flexibly switches between two sub-architectures depending on configuration. One path employs a RepViTBlock with Vision Transformer characteristics, using depthwise convolution as a token mixer to capture long-range dependencies and positional associations between interacting sheep in overlap scenarios. The other path uses a RepBottleneck with multi-branch RepConv structures that fuse three-by-three convolutions, one-by-one convolutions, and identity branches during training, then collapse them into equivalent single convolutions at inference, enhancing fine-grained posture recognition without added cost. A squeeze-and-excitation module within the block adaptively amplifies key channel responses while suppressing background noise from overlapping animals.

Third, an efficient multi-scale attention module in the network’s neck divides feature maps into groups along the channel dimension without reduction, establishing collaborative attention mappings along both horizontal and vertical directions through parallel branches. Two channel attention maps are aggregated through multiplication to achieve cross-channel feature enhancement, while spatial attention weights derived through global average pooling and Softmax operations help the model focus on the key regions associated with mounting behavior. Fourth, the bounding-box regression loss was replaced with a hybrid Inner-SIoU function that combines the angular cost term of SIoU, which guides optimization for low-quality anchors by minimizing angular deviation between predicted and ground-truth boxes, with auxiliary bounding boxes from the Inner-IoU concept that enhance robustness against the drastic scale variations common in crowded pens.

The performance results demonstrate that these four components work synergistically. Trained from scratch on identical hardware, HSM-YOLO achieved a mean Average Precision of 90.2 percent with an F1 score of 88.1 percent, compared to 84.5 percent mAP for the unmodified YOLOv11n baseline. That 5.7 percentage point improvement came alongside a 15.6 percent reduction in computational complexity, down to 5.4 billion floating-point operations, and a model containing just 2.2 million parameters in a 4.5 megabyte package. The system processed 135.9 frames per second on an RTX 4090 graphics card, outpacing every competing model tested, including RT-DETR-l, YOLOv8n, YOLOv10n, and YOLOv12n. Ablation experiments confirmed each module’s contribution: the dual-path C3k2_BiRep added 3.0 points of mAP, the attention mechanism 2.7 points, and the new loss function 2.2 points, while the separable convolutions cut computation by 18.8 percent.

Perhaps most tellingly, the labeling strategy itself proved decisive. When the team compared their three-class annotation against a simpler two-class approach that treated non-sexual mounting as background, the three-class version improved mAP by 2.9 percentage points. The difference was most dramatic for false positives on the mounted class, which plummeted from 954 to 222, a reduction of 76.7 percent, showing that explicitly teaching the model about non-reproductive mounting prevents it from misclassifying irrelevant sheep instances as fertile animals. Across the four imaging scenarios, HSM-YOLO consistently outperformed competitors, achieving 94.6 percent mAP in the favorable side-pen conditions and maintaining 77.3 percent even under the robot’s dynamic viewpoints and motion blur, where rival models degraded severely.

The practical implications extend beyond benchmark numbers. The researchers built a complete monitoring system with a graphical interface supporting real-time detection from single images, local video files, and RTSP IP camera streams, triggering alerts whenever sexual mounting is detected. Testing with 2560-by-1440 resolution video at 25 frames per second showed an average end-to-end detection speed of 47.9 frames per second, comfortably exceeding the playback rate of the source footage. The team also deployed the model on an NVIDIA Jetson Orin NX edge device, where its TensorRT engine ran 7.4 percent smaller and 12.7 percent faster in latency than the baseline YOLOv11, reaching 15.5 end-to-end frames per second.

Generalization testing on 850 images from an entirely different commercial farm in Xiaoshan District, Hangzhou, with distinct lighting, pen layouts, and camera viewpoints, showed HSM-YOLO maintaining 78.8 percent mAP versus 72.6 percent for the baseline, with smaller performance degradation. The researchers acknowledge limitations: the model was trained exclusively on Hu sheep in Zhejiang Province, cannot identify the individual identity of the mounted ewe, lacks distant small-scale samples, and struggles with continuous real-time analysis on memory-constrained edge hardware. Future work will collect video from multiple breeds including Small-tailed Han and Dorper sheep, introduce infrared imaging for all-weather monitoring, integrate tracking and animal re-identification to build a unified detection-tracking-identification framework, and pursue further lightweighting through channel pruning and knowledge distillation. If successful, these advances could transform estrus detection from a labor-intensive guessing game into an automated, round-the-clock precision operation for the sheep farming industry.

Subject of Research: AI Camera System Spots Fertile Ewes in Seconds During Crowded Barn Mounting Events

Article Title: Detection of mounting behavior and mounted ewes in group-housed environments using HSM-YOLO algorithm

Article References: Wang, Y., Gao, M., Hua, Z., Shakeel, A., Ren, C., & Wang, K. (2026). Detection of mounting behavior and mounted ewes in group-housed environments using HSM-YOLO algorithm. Smart Agricultural Technology, 15, Article 102639. https://doi.org/10.1016/j.atech.2026.102639

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102639

Keywords: HSM-YOLO, estrus detection, sheep farming, computer vision, deep learning, YOLOv11n, precision livestock farming, mounting behavior, Hu sheep, agricultural AI, Detection, mounting

The problem the team set out to solve is deceptively hard. Traditional livestock monitoring relies on wearable sensors such as accelerometer collars, GPS tags and RFID ear tags. These devices work, but they stress the animals, break or fall off, run out of battery, and become impractical when dozens of animals share a small space. Camera-based detection avoids all of that, but it introduces its own challenge: sheep in group housing look remarkably alike, cluster tightly, block one another from view, and shift posture constantly. A standard detection model thrown at such a scene tends to miss animals, confuse eating with drinking, and demand more computing power than a farm can afford.

The researchers, publishing in the journal Artificial Intelligence in Agriculture, started from YOLOv11n, the nano version of the popular You Only Look Once object detection family, and rebuilt it in three places. The first change targets raw efficiency. They replaced seven standard convolution layers in the network’s backbone and neck with GSConv, a hybrid module that fuses dense standard convolution with cheap depthwise separable convolution and then shuffles the channels to restore cross-channel communication. The result is a module whose computational cost approaches half that of a standard convolution, cutting the model’s parameters by roughly 12.8 percent and its floating-point operations by 12.7 percent — without hurting accuracy.

The second change is an attention mechanism called LWGA-Lite, a slimmed-down version of a lightweight grouped attention design. Attention modules tell a neural network where to look, and in a sheep barn that matters enormously: the difference between a sheep that is eating and one that is drinking lies in whether its head is at the feed trough or the water valve, while standing versus lying depends on limb geometry and the gap between belly and floor. LWGA-Lite splits features into four parallel paths — gated point attention for fine pixels, local attention for contours, sparse medium-range attention for trunk structure, and sparse global attention for scene-level context — and applies global attention only to the most salient image tokens rather than the whole frame. The lightweight redesign halved attention head dimensions, swapped in depthwise separable convolutions, and replaced costly upsampling operations, trimming parameters while keeping detection quality essentially intact.

The third change concerns how the model learns to draw boxes. The default Complete IoU loss couples width and height penalties together, which can cause slow or oscillating convergence when animals stretch, crouch or overlap. The team substituted the Efficient IoU loss, which decouples width and height into independent penalty terms. The payoff was clearest for the rarest behaviour in the dataset: drinking. Because only a few sheep drink at any moment, drinking samples were scarce, yet the EIoU loss lifted the drinking category’s strict accuracy metric from 0.893 to 0.925 — a gain of 3.2 percentage points on that behaviour alone.

To train and test the system, the researchers built a dataset from a real farm rather than a staged laboratory. Over the winter of 2024 to 2025, an infrared dome camera recorded the pen at 25 frames per second, capturing day and night in equal measure. One thousand raw images containing 13,000 behaviour instances were annotated with fine polygon outlines, then expanded through a two-stage augmentation pipeline — offline rotations, flips, crops and noise, plus online mosaic and copy-paste strategies that deliberately oversampled drinking — into 4,755 training images holding 61,430 instances. Validation and test sets were left untouched to keep the evaluation honest.

The numbers that emerged are striking for a model this small. GLE-YOLOv11n reached a mean average precision of 0.9479 under the strict IoU threshold range and 0.9814 under the looser standard, with precision of 0.9681 and recall of 0.9917. It beat its own YOLOv11n baseline by 1.69 percentage points on the strict metric while carrying only 2.20 million parameters and 5.5 billion floating-point operations per inference. Against heavyweight detectors the gap was enormous: Faster R-CNN trailed by more than 16 percentage points on the strict metric while demanding a hundred GFLOPs and 41 million parameters. Ablation experiments confirmed that the three improvements work synergistically — combining EIoU with LWGA-Lite produced gains larger than the sum of their individual contributions.

Robustness testing pushed the model into the scenarios that break ordinary detectors. Under severe occlusion, where more than 60 percent of a sheep’s body is hidden, mean detection confidence still held at 0.73. In one side-by-side comparison, a heavily occluded sheep that the baseline detected with confidence 0.43 was detected by the improved model at 0.72. The model also generalized better across pens and breeds: when tested on Dorper crossbred sheep it had never seen, its strict accuracy dropped only 2.27 points, compared with 4.86 for the baseline, and in an external farm in Xinjiang it spotted small black lambs the baseline missed entirely.

Then came the test that most published models never face: real deployment. The team compiled the network into a TensorRT FP16 engine and ran it on an NVIDIA Jetson Nano B01, a low-power embedded board. Inference jumped from 6.61 to 15.0 frames per second at 640-by-640 resolution, with accuracy falling only 0.26 percentage points after conversion. The complete system — camera, Jetson Nano, power supply and a fan-cooled protective housing — was bolted above a pen at the commercial farm and left running for thirty days straight. It achieved 100 percent uptime, held a steady 15.0 frames per second, averaged 5.52 watts of input power, and kept its hardware temperature near 63.6 degrees Celsius without a single crash or forced restart, even through feeding, cleaning and shifting light from day to night.

The implications reach beyond sheep. Cloud-based livestock AI raises latency, cost and privacy concerns, and many published models are validated only on gaming-grade GPUs that no farm owns. This work shows the full pipeline — dataset, architecture, loss design and hardware-level optimization — that takes a detection model from a paper to a barn wall. The authors are candid about limits: the system tracks behaviours at group level, not individuals, since it lacks multi-object tracking and re-identification, and fast motion blur remains difficult. Future versions will add lightweight tracking and aggregate detections over time to flag flock-level anomalies, such as a sudden drop in eating that could signal feed problems or disease. For now, the quiet achievement is that a two-hundred-dollar-class computer can watch a flock of sheep, understand what it sees, and never blink.

Subject of Research: A lightweight deep learning model for real-time sheep behaviour recognition on edge computing devices

Article Title: GLE-YOLOv11n: A lightweight network for real-time recognition of sheep behaviours in edge computing environments

Article References: Dong, R., Wei, X., Zheng, D., Liu, Y., Tong, Y., Li, W., Shen, W., & Fu, X. (2026). GLE-YOLOv11n: A lightweight network for real-time recognition of sheep behaviours in edge computing environments. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.09.007

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.09.007

Keywords: sheep behaviour recognition, edge computing, YOLOv11n, GSConv, attention mechanism, EIoU loss, Jetson Nano, TensorRT, precision livestock farming, computer vision, animal welfare, model lightweighting