The duck egg industry has always dominated an important and indispensable position in the agricultural field. However, the huge amount of egg duck breeding has also brought some challenges to the current industry in recent years. Among them, the low-level automation is ever lagging behind to limit the production efficiency and product quality in duck and egg breeding. Furthermore, the duck eggs can be collected one by one by hand, leading to low work efficiency and high labor intensity. Manual operation cannot fully meet the large-scale production in modern agriculture. At the same time, the frequent entry and exit of breeders into duck farms has inevitably brought some hidden dangers to the health of ducks, such as various pathogens and pollutants from the external environment. This research aims to accurately identify and position the free-range duck eggs at medium and long distances in the key technology of egg collection robots. The automation and intelligence of duck egg collection were then realized using the YOLOv8n network model. A P2 detection head was also added to detect the small target. Accurate detection was achieved in the large-size feature maps that retained the more target features. The CBAM attention module was inserted at the end of the backbone network. The attention was weighted in the channel and spatial dimensions. The neural network was effectively enhanced to focus on the important features, and then suppress the interference of unnecessary features. At the same time, the Involution operator was introduced to significantly improve the detection performance for the duck eggs at medium and long distances. The traditional convolution was then modified in the channel-to-channel information exchange and receptive field. A visual positioning was proposed using ROS. The high-precision and real-time visual positioning of target duck eggs was achieved using machine vision and dynamic TF transformation when the camera was used in the moving state with the robot. The test results show that the weight of the improved model (YOLOv8n-P2-CBAM-INV) was 20.8 MB, and the floating-point operations per second were 51.4 G. The accuracy and recall rates of the improved model were 98.30% and 96.60%, respectively. The mean average precision on the test set increase 7.1 percentage points compared with the original YOLOv8n. The mean average precision of this improved model reached 98% at a distance of 6.0 m, which was increased 29.4 percentage points compared with the original model. The visual positioning test of the mobile robot was carried out to verify the range of 0.5 to 3.0 m. The high-precision and real-time motion positioning was realized at the medium and long-distance duck eggs by the mobile egg-picking robot in the laboratory. The positioning error was controlled within 0.03 m. This improved model can be deployed to the mobile edge platform of image processing. The finding can also provide support to the mobile egg-picking robots, together with the similar autonomous mobile harvesting robots.

To accurately and efficiently detect dead caged laying ducks, thereby reducing reliance on manual inspection, this study proposes a method that integrates infrared thermography with deep learning technology. A lightweight object detection algorithm is developed, utilizing YOLO v8n as the baseline model. The backbone network is replaced with StarNet, which is based on “Star Operate”. Additionally, the C2f-Star structure is designed by combining the Star Block from StarNet with the C2f module, and it is inserted into the Neck structure of the baseline model. Lightweight module L-SPPF replaces the SPPF module in the baseline model to enhance feature augmentation. Furthermore, a lightweight shared convolutional detection head, termed SCSB-Head, is introduced to reduce computational complexity. These improvements collectively form a lightweight object detection algorithm named SLSS-YOLO. Experimental results show that SLSS-YOLO achieves mAP@50%-95%, precision, and recall scores of 80.50%, 99.44%, and 98.46%, respectively. Compared to the baseline model, these metrics improve by 1%, 1.98%, and 0.26%, respectively. In terms of model size and detection speed, SLSS-YOLO has 1.44 M parameters and 4.6 G FLOPs, achieving an FPS rate of 134.9 f/s. This represents a reduction of 52.16% and 43.90% in parameters and FLOPs, respectively, while increasing FPS by 5.4 f/s compared to the baseline model. Moreover, an object tracking model is constructed using SLSS-YOLO and Hybrid-SORT. Tracking tests demonstrate that Hybrid-SORT achieves zero ID-Switches, with a detection speed of 10.9 ms/f. It outperforms Bot-SORT, ByteTrack, Deep OC-SORT, and OC-SORT in terms of tracking performance. Therefore, the proposed thermal infrared detection method can effectively identify and track dead ducks in complex cage environments, providing a reference for automated inspection in caged duck farms.