Behavioral information of beef cattle has been widely acknowledged as the critical indicator related to physiological health status, welfare, and production. Therefore, it is often required to accurately recognize cattle behaviors for health evaluation and early abnormality warning in precision livestock farming. However, visual monitoring has been severely restricted under nighttime breeding environments due to the low-illumination imaging. Insufficient lighting, uneven illumination distribution, and motion-induced blur have significantly degraded image quality, thus resulting in blurred appearances and weak behavioral feature representations. Existing machine vision can also suffer from frequent missed detections and false positions during feature extraction and classification. In this study, a behavior recognition framework, named Night-DETR, was proposed to specifically monitor beef cattle under complex nighttime scenarios. First, a low-light image enhancement with SCINet was introduced as a preprocessing module to improve the visual quality of nighttime monitoring images. Image brightness and structural clarity were effectively increased by suppressing noise after enhancement, thereby providing more informative and stable visual inputs for downstream behavior recognition. Second, the baseline RT-DETR architecture was systematically redesigned to improve the detection accuracy and computational efficiency. Particularly, StarNet was adopted as the backbone network to replace conventional heavy feature extractors. Multi-scale features were efficiently represented to reduce parameter redundancy and computation. An Adaptive Interactive Feature Integration module with an Efficient Discriminative Frequency Domain-based FFN (AIFI-EDFFN) was designed to treat feature ambiguity and inter-object interference under low-light conditions. Behavioral features were enhanced to suppress irrelevant background responses to contextual interactions among neighboring cattle, thereby strengthening contour, posture, and spatial relationship representations. In addition, an Intensity Enhance Layer Cross-scale Feature Fusion module (IELC3) was constructed to optimize multi-scale feature aggregation. An illumination enhancement layer was embedded into the cross-scale fusion. Limb structures, body orientation, and posture features were typically obscured to emphasize them in nighttime environments, thus improving the sensitivity to subtle behavioral changes under weak lighting. A series of experiments was conducted to evaluate the Night-DETR model. The performance of Night-DETR was compared with the representative and state-of-the-art object detection frameworks, including Faster R-CNN, TOOD, FCOS, YOLOv11n, YOLOv12n, YOLOv13n, and the original RT-DETR. Experimental results demonstrated that Night-DETR consistently performed best in nighttime detection scenarios. Compared with the baseline RT-DETR model, Night-DETR improved precision, recall, and mean average precision by 5.3, 5.8, and 5.1 percentage points, respectively, thus achieving 93.4% precision, 90.8% recall, and 91.3% mAP@0.5. Moreover, the computational complexity was significantly reduced, with the number of parameters and floating-point operations decreased by 55.8% and 58.2%, respectively. A lightweight architecture was obtained with only 8.8 M and 23.8 G. Cross-scene transfer learning experiments were conducted to verify the generalization of the improved model among different farming scenarios. An average precision of 90.8% was achieved after transfer learning, which was 1.4 percentage points higher than the original. Night-DETR shared strong cross-scene generalization in various farming environments. Additionally, the robustness of the model was evaluated under extremely low-light conditions. The performance curves were plotted for precision, recall, and mAP@50 among different illumination gradients. Night-DETR maintained stable performance even in severely dark scenarios, indicating its strong robustness under extremely low-light environments. Overall, the Night-DETR model can be expected to provide accurate behavioral perception under low-illumination environments with the lightweight deployment in the intelligent nighttime health monitoring and abnormal behavior early-warning systems. The findings can also offer strong technical support for precision livestock farming in breeding environments.

Cattle monitoring can rely mainly on manual inspections in the fence farming. Once the cattle escaped near the edge of the fence, the conventional regulatory system could not fully detect and intervene in time, resulting in the loss of cattle. In this study, a cattle area monitoring and boundary-crossing early warning system was proposed for the cattle using fence-type breeding. Instance segmentation and multi-object tracking, named STWBD (Segment Tracking with Boundary Detection), were also utilized to monitor the cattle activities in real time within the fenced area. Intelligent early warnings of the potential boundary-crossing behaviors were then provided to enhance the safety level of the fenced cattle farming. Firstly, an improved RDL-YOLO11n-seg instance segmentation was constructed into the framework. The C3K2-RetBlock module was also introduced to the original YOLO11n-seg, in order to enhance the response capability to the spatial displacement of the cattle. The C2DA structure was combined to improve the expression of the multi-scale features. The Segment-LSCD module was employed to optimize the segmentation head. The object detection and instance segmentation were facilitated to accurately capture the cattle contours and postures. As a result, the spatial perception and real-time monitoring of the cattle were enhanced within the fenced area. Secondly, the DeepOCSORT algorithm was assigned an identity ID to each cattle, particularly for the continuous tracking in different scenarios. Some challenges were effectively handled, such as the small target scale, severe occlusion, low lighting, and cattle close to the boundary. The continuity and stability of the multi-object tracking were provided for the reliable trajectory data to detect the later boundary-crossing behavior. In terms of the fence boundary design, the actual structure of the cattle shed fences was combined to construct a polygonal electronic fence. The safe activity area of the cattle was defined for the continuous analysis of the movement trajectories and position distribution of the cattle. The system was integrated with some functions, such as the trajectory visualization, real-time speed tracking, and statistics of the cattle population. The STWBD was used to monitor the boundary-crossing behavior of the cattle in real time and promptly trigger the intelligent early warnings. Experimental results showed that the precision, recall, and segmentation accuracy of the RDL-YOLO11n-seg model reached 94.9%, 87.4%, and 92.5%, respectively, which were 2.0, 3.1, and 3.1 percentage points higher than the original, and the parameters were reduced by 17.8%, with a floating-point operation count of 8.6 G. The high-order tracking accuracy (HOTA), multiple object tracking accuracy (MOTA), and identity F1-Score (IDFI) of DeepOCSORT were 81.4%, 91.6%, and 95.8%, respectively, while the precision, recall, and F1 score of the escape event detection were 96.2 %, 100.0 %, and 98.1 %, respectively. The movement trajectories of the cattle were efficiently monitored in the small-scale fenced farming environments. The boundary-crossing escape events were effectively identified to provide the reliable technical support for the intelligent early warning and supervision of the cattle. This finding can provide real-time access to the cattle population, distribution, and movement patterns. While the risk was predicted after trajectory and speed analysis. The finding can offer a scientific basis for the secure and efficient management of the confined cattle farming.