Deep transfer learning (DTL), which incorporates new ideas from deep neural networks into transfer learning (TL), has achieved excellent success in computer vision, text classification, behavior recognition, and natural language processing. As a branch of machine learning, DTL applies end-to-end learning to overcome the drawback of traditional machine learning that regards each dataset individually. Although some valuable and impressive general surveys exist on TL, special attention and recent advances in DTL are lacking. In this survey, we first review more than 50 representative approaches of DTL in the last decade and systematically summarize them into four categories. In particular, we further divide each category into subcategories according to models, functions, and operation objects. In addition, we discuss recent advances in TL in other fields and unsupervised TL. Finally, we provide some possible and exciting future research directions.

The availability of autonomous orchard robots could alleviate the conflict caused by rising labor costs and labor shortages. The essential technical requirements are autonomous localization and mapping which rely on place recognition to explore data associations. This letter presents a novel LiDAR-based place recognition algorithm for unstructured and large-scale orchards. Concretely, we propose a discriminative global representation, spatial binary pattern (SBP), that encodes three-dimensional (3D) spatial distributed scan into an eight-bit binary pattern. In addition, an efficient two-stage hierarchical re-identification process is proposed. The attention score map is introduced for task-relevant features and preliminary candidates retrieval. The overlap re-identification is used to align a pair of descriptors to confirm the final loop closure index. Experiments on orchard and public datasets have been conducted to evaluate the performance of the proposed method, our method achieves a higher recall rate and localization accuracy. Moreover, experiments on the long-term outdoor dataset KITTI further demonstrate the generality.

The fusion of light detection and ranging (LiDAR) and inertial measurement unit (IMU) sensing information can effectively improve the environment modeling and localization accuracy of navigation systems. To realize the spatiotemporal unification of data collected by the IMU and the LiDAR, a two-step spatiotemporal calibration method combining coarse and fine is proposed. The method mainly includes two aspects: (1) Modeling continuous-time trajectories of IMU attitude motion using B-spline basis functions; the motion of the LiDAR is estimated by using the normal distributions transform (NDT) point cloud registration algorithm, taking the Hausdorff distance between the local trajectories as the cost function and combining it with the hand–eye calibration method to solve the initial value of the spatiotemporal relationship between the two sensors’ coordinate systems, and then using the measurement data of the IMU to correct the LiDAR distortion. (2) According to the IMU preintegration, and the point, line, and plane features of the lidar point cloud, the corresponding nonlinear optimization objective function is constructed. Combined with the corrected LiDAR data and the initial value of the spatiotemporal calibration of the coordinate systems, the target is optimized under the nonlinear graph optimization framework. The rationality, accuracy, and robustness of the proposed algorithm are verified by simulation analysis and actual test experiments. The results show that the accuracy of the proposed algorithm in the spatial coordinate system relationship calibration was better than 0.08° (3δ) and 5 mm (3δ), respectively, and the time deviation calibration accuracy was better than 0.1 ms and had strong environmental adaptability. This can meet the high-precision calibration requirements of multisensor spatiotemporal parameters of field robot navigation systems.

Orchard robots tasks rely on highly accurate, real-time trajectory estimation and map-building. In this paper, a novel tightly-coupled LIDAR inertial simultaneous localization and mapping (SLAM) system, SG-ISBP-SLAM, is proposed. The algorithm involves both ground optimization and loop closure detection. Aiming at the uneven orchard ground, this paper designs a ground segmentation method from a global perspective. It takes the nearest neighbor seed plane as the base-line and iteratively grows a global plane based on principal component analysis (PCA). The LIDAR scan is divided into 3D concentric zone representations to assign an appropriate density of cloud points among bins. Based on the partition strategy, the improved spatial binary pattern (ISBP) is encoded for lower time-consuming loop closure detection. To validate the performance of the proposed algorithm, qualitative and quantitative experiments have been conducted. Experimental results indicate that SG-ISBP-SLAM provides low-time consumption and reliable non-flat ground segmentation capabilities. Moreover, the loop module can efficiently correct robot localization trajectory drift.

High accuracy of locating the sphere center imaging points is crucial for the vision measurement system. This paper investigates the effect of projection projective model on extraction of the sphere center, upon which the sub-pixel edge-location algorithm based on the sphere projection model is proposed. The sphere center projection model is established by analysing the process of sphere imaging, and the sub-pixel edge-location algorithm consisted of the novel edge model and the improved Zernike moment computing is studied. The novel edge model based on the Erf (error function) is adopted for modelling the practical edge part, which obtains the edge distribution points with high precision. Then a closed-form solution of edge locating error compensation is calculated based on Zernike moments. Finally, the proposed method is verified via simulations and experiments. The relevant results show that the proposed method can accurately locate the sphere center imaging point, which further improves the precision and robustness of the vision measurement system.

Process monitoring (PM) is important for improving product quality and ensuring plant safety in industrial systems. Recently, canonical correlation analysis (CCA)-based PM has shown excellent performance. The core idea is to first seek the relationship between two sets of process variables via CCA, then construct a residual generator and determine test statistics, and finally achieve the best monitoring. However, the global and local structure of process variables is not fully exploited. In this article, we propose a new nonlinear PM approach called JSKCCA-GL by incorporating joint sparse kernel CCA and graph priors. Technically, imposing the ℓ2,0-norm joint sparse constraints can capture the global row-wise structure, and embedding the graph Laplacian priors can preserve the local structured variable correlation information. Further, an efficient alternating minimization algorithm is developed by integrating least squares and iterative hard thresholding. The superiority of the proposed JSKCCA-GL is extensively verified on the benchmark Tennessee Eastman process (TEP) and a practical cylinder-piston assembly (CPA) of marine diesel engines. In particular, the fault detection rate (FDR) values for unknown-type faults in the TEP have been improved by around 16.10% on average.