Robot perception difficulties in complex environments seriously affect robot operational efficiency. The application of generative big model (GBM) technology on robots through the Agriculture Internet of Things (AIoT) can solve the difficulty of low-operational efficiency. Therefore, this article proposes an automatic decision-making localization algorithm for robots fused with GBMs in AIoT. First, in the AIoT, the complex agricultural scene information is acquired in real-time by the Realsense D435i equipped on the robot, which is accessed in the GBM through a proprietary network to realize real-time intelligent sensing and control between the environmental information and the robot. Then, the innovative reinforcement learning method based on human solid feedback (S-RLHF) and the automatic generation method of weakly supervised fine-tuning data (WS-FT DAGM) are designed in the generative large model. At the same time, by combining the characteristics of the robot operation, two generative significant model recommendation methods are designed, which solves the problem of the difficulty of the target perception in the complex agricultural scene. Finally, by integrating the AIoT and generative large models, the critical method of real-time analysis of crop shading characteristics by the large model is innovatively proposed to solve the problem of low efficiency in robot operation. In the robot test experiments, the operation efficiency using the fusion of AIoT and generative large model reaches more than 92%, significantly improving the operation efficiency compared to the small model (CNN) method without AIoT and generative large model in the traditional agricultural robot.

Unstructured grid generation is more suitable for finishing the joint inversion of gravity and magnetic anomalies in the case of undulating observation surfaces and irregular field sources. To realize the joint inversion with unstructured grid of gravity and magnetic anomalies corrupted by remanence, we proposed a joint inversion method of gravity and magnetic analytic signal (MAS) with unstructured grid, and an adaptive subdivision method based on anomaly feature is introduced to improve the computation efficiency. The least-squares method was used to perform the computation of cross-gradient item in joint inversion for non-homologous gravity and magnetic source. Synthetic tests on a dipping-slab model show that the proposed method can complete the joint inversion, and can effectively recover the distribution characteristics of the field source without precision loss, and can reduce calculation time and storage space. Finally, we applied our new method to real gravity and magnetic data to ascertain the distribution of magnetite in Shandong Province, China. We obtained the range of favorably mineralized regions based on the inversion results, which provides an important basis for further exploration.

Cross-gradient joint inversion of gravity and seismic data can more accurately provide the characteristics of subsurface density and velocity structure. The subsurface is usually divided into triangular grids to accurately simulate the undulating terrain and irregularity of geological bodies in the inversion, and the existing cross-gradient inversion with triangular grid is finished by the linear trend method. However, the density and velocity changes are not all linear. To better describe both nonlinear and linear variation feature of the physical property, we propose the second-order finite-difference cross-gradient joint inversion method of gravity and seismic data with triangular grid, which can obtain high-resolution results and effectively reflect the linear and nonlinear physical property changes. We also compare the effect of cross-gradient inversion results computed by different order finite-difference method and verify that the second-order finite-difference method is more reasonable according to the computational efficiency and accuracy. To reveal the distribution of the polymetallic minerals in the Lu (Lujiang)-Zong (Zongyang) ore concentration area, we carry out regional gravity and profile seismic measurements in this area. We first obtain the 2-D density and velocity results, and then use the 2-D density as a constraint to compute the 3-D density distribution. The results reveal the distribution of six minerals and that the burial depth of high-density polymetallic ores range from 567 to 959 m, which provides reliable information for subsequent exploitation.

In the 3-D density inversion calculation of gravity data, the entire subsurface space is discretized into rectangular prisms, and the density value of each prism is calculated. However, the method of dividing the entire space often results in invalid calculations. In this study, we proposed a highly efficient density inversion method using a locally adaptive unstructured mesh. In this method, the inversion scope is reduced by using the feature that the zero value of the tilt angle method corresponds to the edge of the field source. In addition, the local inversion area is divided by unstructured meshes, which can better represent irregular geological bodies and undulating terrain. In the proposed method, the size of the grid cells is changed according to the value of the tilt angle to reduce the amount of mesh, which reduces the computational complexity and improves the computational efficiency. In addition, by introducing a volume weighting function, the sensitivity of grid cells of different sizes can be balanced. Through synthetic modeling experiments, we verified that the locally adaptive unstructured mesh method can improve the efficiency and accuracy of inversion, flexibly deal with undulating terrain, and obtain the distribution features of irregular bodies. We applied this method to gravimetric data of the North Qinling region of the Shaanxi province, which clearly shows the spatial distribution information of five high-density geological bodies, and the depth range of ore bodies formed is from 240 to 1481 m according to our inversion results.

Ground segmentation is an important preprocessing task for autonomous vehicles (AVs) with 3D LiDARs. However, the existing ground segmentation methods are very difficult to balance accuracy and computational complexity. This paper proposes a fast point cloud ground segmentation approach based on a coarse-to-fine Markov random field (MRF) method. The method uses the coarse segmentation result of an improved local feature extraction algorithm instead of prior knowledge to initialize an MRF model. It provides an initial value for the fine segmentation and dramatically reduces the computational complexity. The graph cut method is then used to minimize the proposed model to achieve fine segmentation. Experiments on two public datasets and field tests show that our approach is more accurate than both methods based on features and MRF and faster than graph-based methods. It can process Velodyne HDL-64E data frames in real-time (24.86 ms, on average) with only one thread of the I7-8700 CPU. Compared with methods based on deep learning, it has better environmental adaptability.

Effective semantic segmentation of lane marking is crucial for construction of high-precision lane level maps. In recent years, a number of different methods for semantic segmentation of images have been proposed. These methods concentrate mainly on analysis of camera images, due to limitations with the sensor itself, and thus far, the accurate three-dimensional spatial position of the lane marking could not be obtained, which hinders lane level map construction.This article proposes a lane marking semantic segmentation method based on LIDAR and camera image fusion using a deep neural network. In the approach, the object of the semantic segmentation is a bird’s-eye view converted from a LIDAR points cloud instead of an image captured by a camera. First, the DeepLabV3+ network image segmentation method is used to segment the image captured by the camera, and the segmentation result is then merged with the point clouds collected by the LIDAR as the input of the proposed network. A long short-term memory (LSTM) structure is added to the neural network to assist the network in semantic segmentation of lane markings by enabling use of time series information. Experiments on datasets containing more than 14,000 images, which were manually labeled and expanded, showed that the proposed method provides accurate semantic segmentation of the bird’s-eye view LIDAR points cloud. Consequently, automation of high-precision map construction can be significantly improved. Our code is available athttps://github.com/rolandying/FusionLane.

Moving vehicle detection in dynamical scene is a significant but challenging problem in these days. A new and effective approach to extract moving vehicles is proposed in this paper. In our method, Harris corner and Lucas-Kanade (L-K) optical flow was adopted to generalize feature-point optical flow field between two consecutive frames which obtained from monocular moving camera, and then vector quantization (VQ) was used to cluster the optical flow field using similarity measurement of Euclidean distance and similarity coefficient. At last, through calculating the variance for each class to eliminate the mismatched optical flow and extract the vehicles from background. Experiment result shown that our method has an excellent performance in eliminating the mismatched optical flows. It can extract vehicles from dynamical scene exactly and can also meet the real-time requirement. Furthermore, it provides accurate information for next step of vehicle tracking.