As global pork consumption rises, livestock farms increasingly adopt deep learning-based automated monitoring systems for efficient pigsty management. Typically, a system applies a pre-trained model on a source domain to a target domain. However, real pigsty environments differ significantly from existing public datasets regarding lighting conditions, camera angles, and animal density. These discrepancies result in a substantial domain shift, leading to severe performance degradation. Additionally, due to variations in the structure of pigsties, pig breeds, and sizes across farms, it is practically challenging to develop a single generalized model that can be applied to all environments. Overcoming this limitation through large-scale labeling presents considerable burdens in terms of time and cost. To address the degradation issue, this study proposes a self-training-based domain adaptation method that utilizes a single label on target (SLOT) sample from the target domain, a genetic algorithm (GA)-based data augmentation search (DAS) designed explicitly for SLOT data to optimize the augmentation parameters, and a super-low-threshold strategy to include low-confidence-scored pseudo-labels during self-training. The proposed system consists of the following three modules: (1) data collection module; (2) preprocessing module that selects key frames and extracts SLOT data; and (3) domain-adaptive pig detection module that applies DAS to SLOT data to generate optimized augmented data, which are used to train the base model. Then, the trained base model is improved through self-training, where a super-low threshold is applied to filter pseudo-labels. The experimental results show that the proposed system significantly improved the average precision (AP) from 36.86 to 90.62 under domain shift conditions, which achieved a performance close to fully supervised learning while relying solely on SLOT data. The proposed system maintained a robust detection performance across various pig-farming environments and demonstrated stable performance under domain shift conditions, validating its feasibility for real-world applications.
With the increasing demand for pork, improving pig health and welfare management productivity has become a priority. However, it is impractical for humans to manually monitor all pigsties in commercial-scale pig farms, highlighting the need for automated health monitoring systems. In such systems, object detection is essential. However, challenges such as insufficient training data, low computational performance, and generalization issues in diverse environments make achieving high accuracy in unseen environments difficult. Conventional RGB-based object detection models face performance limitations due to brightness similarity between objects and backgrounds, new facility installations, and varying lighting conditions. To address these challenges, this study proposes a DOG (Depth-Oriented Gray) image generation method using various foundation models (SAM, LaMa, Depth Anything). Without additional sensors or retraining, the proposed method utilizes depth information from the testing environment to distinguish between foreground and background, generating depth background images and establishing an approach to define the Region of Interest (RoI) and Region of Uninterest (RoU). By converting RGB input images into the HSV color space and combining HSV-Value, inverted HSV-Saturation, and the generated depth background images, DOG images are created to enhance foreground object features while effectively suppressing background information. Experimental results using low-cost CPU and GPU systems demonstrated that DOG images improved detection accuracy (AP50) by up to 6.4% compared to conventional gray images. Moreover, DOG image generation achieved real-time processing speeds, taking 3.6 ms on a CPU, approximately 53.8 times faster than the GPU-based depth image generation time of Depth Anything, which requires 193.7 ms.
Due to the global population increase and the recovery of agricultural demand after the COVID-19 pandemic, the importance of agricultural automation and autonomous agricultural vehicles is growing. Fallen person detection is critical to preventing fatal accidents during autonomous agricultural vehicle operations. However, there is a challenge due to the relatively limited dataset for fallen persons in off-road environments compared to on-road pedestrian datasets. To enhance the generalization performance of fallen person detection off-road using object detection technology, data augmentation is necessary. This paper proposes a data augmentation technique called Automated Region of Interest Copy-Paste (ARCP) to address the issue of data scarcity. The technique involves copying real fallen person objects obtained from public source datasets and then pasting the objects onto a background off-road dataset. Segmentation annotations for these objects are generated using YOLOv8x-seg and Grounded-Segment-Anything, respectively. The proposed algorithm is then applied to automatically produce augmented data based on the generated segmentation annotations. The technique encompasses segmentation annotation generation, Intersection over Union-based segment setting, and Region of Interest configuration. When the ARCP technique is applied, significant improvements in detection accuracy are observed for two state-of-the-art object detectors: anchor-based YOLOv7x and anchor-free YOLOv8x, showing an increase of 17.8% (from 77.8% to 95.6%) and 12.4% (from 83.8% to 96.2%), respectively. This suggests high applicability for addressing the challenges of limited datasets in off-road environments and is expected to have a significant impact on the advancement of object detection technology in the agricultural industry.
Standard Deep Learning-based detectors generally face a trade-off between accuracy and latency, as well as a significant performance degradation when applied to unseen environments. To address these challenges, this study proposes a method that enhances both accuracy and latency by leveraging the static characteristics of fixed-camera pig pen monitoring. Specifically, we utilize background and infrastructure information obtained through a one-time preprocessing step upon camera installation. By integrating this information, we introduce three distinct modules, Background-suppressed Image Generator (BIG), Facility Image Generator (FIG), and Background Suppression Integration (BSI), that improve detection accuracy and operational efficiency without the need for model retraining. BIG creates background-suppressed images that integrate foreground and background information. FIG creates facility mask images that can be used to identify pigs that are occluded by facilities, enabling more efficient learning in unseen environments. BSI leverages both the input image and the background-suppressed image generated by BIG, feeding them into a 3D convolution layer for efficient feature fusion. This difference-aware fusion helps the model focus on foreground information and gradually reduce the domain gap. After training on the German pig dataset and testing on the unseen Korean Hadong pig dataset, the proposed method could improve AP50 accuracy (from 75% to 86%) and Jetson Orin Nano latency (from 67 ms to 41 ms) compared to the baseline model YOLOV12m.