To track the pose of local optical scanners (LOSs) under occlusion and reflection, a robust speckle-textured spherical target and corresponding pose estimation methods are developed in this work, applicable to systems, such as structured light. The proposed method comprises four main stages: first, rigidly attaching the spherical target to a LOS, followed by capturing binocular images and calculating the position of the target; second, generating two panoramic images of the spherical target surface. Third, determining the target's attitude using correlation coefficient analysis matching to panoramic images; and fourth, comprehensive pose determination of the target. To validate the effectiveness of the proposed method under different occlusion and local reflection conditions, the spherical target is rigidly connected to a high-precision six-degree-of-freedom robotic arm in this work. The robotic arm serves a dual purpose: acting as a LOS whose pose is tracked and providing ground truth pose data for accuracy validation. The experimental results demonstrate that the proposed method achieves comparable accuracy under 15% occlusion conditions to state-of-the-art methods under occlusion-free scenarios, providing a valuable reference for existing target-based pose measurement techniques.

Stereo matching is a process to find corresponding points in two images for recovering scene depth. Most prior work relies on epipolar rectification to reduce the correspondence search to one dimension. However, in real-world deployments, camera extrinsics may drift due to shocks, sustained vibration, and thermal effects, thereby introducing non-negligible vertical disparity. Moreover, rectification resampling alters the intrinsic statistics of noise and texture, potentially reducing matching reliability. To address these issues, we propose Unrectified-Stereo, a new paradigm that jointly estimates horizontal and vertical disparities without epipolar rectification. Guided by left and right monocular depth priors, we construct a depth-augmented correlation volume and decompose it into horizontal and vertical cost volumes. A dual-branch iterative framework then performs mutual refinement: confidence-guided stereo cues adaptively select reliable evidence to correct the scale and bias of monocular depth, while the refined depth in turn constrains and updates the two-dimensional disparities. This closed loop progressively transforms coarse, object-level depth priors into pixel-wise, geometrically consistent estimates. We further construct and release unrectified-stereo datasets derived from the Scene Flow, ETH3D, KITTI 2012, and KITTI 2015 datasets for systematic evaluation, featuring vertical disparities up to 22 pixels. Experimental results show that Unrectified-Stereo achieves the best overall performance among the evaluated methods on the constructed unrectified benchmarks, improves robustness in challenging regions, and shows competitive zero-shot transferability across datasets. The source code and data generation scripts are publicly available at https://github.com/zxc04152417/Unrectified-Stereo.

Learning dense motion fields in the bird’s-eye view (BEV) space through self-supervision has become an important research direction in autonomous-driving perception. However, existing methods mainly rely either on spatiotemporal consistency within BEV point cloud sequences or on additional image-based cues such as optical flow, and are therefore constrained by either the limited geometric information provided by a single BEV representation or the need for additional sensors. To address these limitations, we propose Self-CVMotion, a LiDAR-only cross-view self-supervised motion prediction framework that exploits the complementary geometric representations of BEV and range view (RV) to construct supervision signals. Specifically, we design a range-residual-guided spatiotemporal encoding branch in RV, incorporate a normal-consistency term into the optimal transport cost, and construct BEV motion pseudo-labels together with their corresponding RV motion region pseudo-labels from the matching results. Meanwhile, we exploit the geometric correspondence between the azimuth change induced by object motion in BEV and the resulting column shift in RV to formulate a cross-view motion consistency constraint. Experiments on public datasets demonstrate that Self-CVMotion achieves lower prediction errors than existing self-supervised baselines and further narrows the performance gap with fully supervised methods.

High-accuracy and high-efficiency target positioning method is crucial for the 3-D measurement performance of large-scale components. However, most existing methods typically involve the time-consuming process of identification and matching of target markers, which can lead to the measurement system failing to meet the real-time demands of the manufacturing industry. To solve this problem, we propose an end-to-end target monocular positioning method. Specifically, we first design a speckle texture spherical target to eliminate the necessity of identifying and matching target markers. Next, we propose a fast target positioning regression network to model the mapping relationship between the speckle texture information of the spherical target and its spatial position parameters, thereby ensuring the accuracy and efficiency of the target positioning. Experimental results demonstrate that our proposed method achieves a translation error of less than 0.15 mm and a rotation error of less than 0.17(degrees), with a visual positioning time of 0.02 s. In addition, our proposed method can achieve precise target positioning within a measurement range of 2 m. Furthermore, extensive experiments demonstrate the robustness and reliability of our proposed method under various conditions.

Low-light image enhancement is fundamentally constrained by unanchored mappings within perceptually coupled color spaces. In standard sRGB pipelines, joint luminance adjustment and noise suppression can perturb chromaticity, while Retinex-based decomposition lacks a physical definition of target illumination, resulting in an unconstrained solution space and inconsistent spectral fidelity. To address these limitations, we propose CATFormer as a chromatic adaptation transform followed by perceptually decoupled restoration. CAT-Former first uses Dynamic White Point Adaptation (DPA) to estimate a per-pixel Chromatic Adaptation Transform, providing a learnable von Kries implementation that sets the signal’s color representation to the standard white point. The adapted signal is then converted into CIELAB, where the Dual-branch Decoupled Restoration (DDR) network provides the restoration geometry for separately modeling luminance and chrominance degradations under the estimated CAT prior, while avoiding interference. Anchor-based Kernel Attention (AKA) serves as an efficient global-context operator for the adaptation and restoration stages with linear complexity. Extensive experiments on multiple public benchmarks demonstrate that CATFormer achieves state-of-the-art performance. The source code is available at https://github.com/wuky2000/CATFormer.

Depth completion, the task of reconstructing dense depth maps from sparse measurements, is crucial for scene understanding and autonomous systems. Leveraging aligned, high-resolution RGB images as guidance is a common and powerful approach, yet the inherent frequency heterogeneity between RGB and sparse depth data presents a significant challenge for effective cross-modal fusion. Conventional methods often employ simplistic fusion strategies that overlook these distinct frequency characteristics, limiting their ability to fully exploit the complementary nature of RGB and depth information. In this paper, we introduce DIFNet: a Dual-Information Fusion Network, based on a novel frequency-aware fusion paradigm focused on image-guided completion. The core of DIFNet is the Dual Stream Modeling (DSM) block, which explicitly decouples and processes high-frequency edge details and low-frequency smooth regions with tailored architectures, leveraging a spatially-aware Mamba architecture for high-frequency streams and densely connected convolutions for low-frequency streams. Furthermore, DIFNet incorporates an innovative Initial Feature Fusion (IFF) layer to facilitate synergistic multi-scale RGB and depth feature integration from the input stage. Extensive evaluations on KITTI and NYUv2 datasets demonstrate that DIFNet achieves state-of-the-art performance in depth completion with competitive computational efficiency, highlighting the efficacy of our frequency-aware dual information fusion strategy. The code for this work is publicly available at https://github.com/wuky2000/DIFNet.

Traditional line-buffer architectures for stereo rectification require on-chip memory that scales linearly with image width and vertical distortion, making UHD deployment on FPGAs impractical. This brief presents a tile-based architecture in which lens-distortion-induced irregular source footprints are precomputed offline into minimal bounding boxes and fetched via SG-DMA descriptor chains, while per-pixel coordinates are recovered at runtime from sparse grids. By decoupling on-chip memory from vertical distortion, the proposed design reduces BRAM consumption by over 90% relative to equivalent line-buffer designs, enabling UHD stereo rectification on FPGAs for the first time. Implemented on ZynqMP SoC and validated using a PL pattern generator, it processes 4K stereo streams at 30 fps (scalable to 67.1 fps) with an end-to-end latency of 4.38 ms and a total on-chip power of 3.7W, while sustaining 8K streams at 20 fps.

Due to the complexity of the unstructured environment and the high-level requirement of smoothness when a tracked transportation vehicle is traveling, making the vehicle travel as safely and smoothly as when a skilled operator is maneuvering the vehicle is a critical issue worth studying. To this end, this study proposes a trajectory-planning method for human-like maneuvering. First, several field equipment operators are invited to manipulate the model vehicle for obstacle avoidance driving in an outdoor scene with densely distributed obstacles, and the manipulation data are collected. Then, in terms of the lateral displacement, by comparing the similarity between the data as well as the curvature change degree, the data with better smoothness are screened for processing, and a dataset of human manipulation behaviors is established for the training and testing of the trajectory-planning network. Then, using the dynamic parameters as constraints, a two-stage planning approach utilizes a modified deep network model to map trajectory points at multiple future time steps through the relationship between the spatial environment and the time series. Finally, after the experimental test and analysis with multiple methods, the root-mean-square-error and the mean-average-error indexes between the planned trajectory and the actual trajectory, as well as the trajectory-fitting situation, reveal that this study’s method is capable of planning long-step trajectory points in line with human manipulation habits, and the standard deviation of the angular acceleration and the curvature of the planned trajectory show that the trajectory planned using this study’s method has a satisfactory smoothness.

Reducing energy consumption of wheeled robots in urban inspection and unstructured environments is a pressing challenge. This study proposes a human-like trajectory planning method based on deep learning to address energy inefficiency. A convolutional neural network (CNN) with multi-dimensional attention extracts spatial features from driving scenes and radar maps of hazardous areas. Temporal dependencies are captured using an improved long short-term memory (LSTM), where state information is added to the gate update module. Power, speed, and angular velocity are incorporated as constraints to enhance trajectory mapping accuracy. Experimental results show that, compared with traditional and state-of-the-art methods, the proposed approach significantly reduces cumulative power consumption and improves accuracy in predicting future trajectories. The model effectively learns human manipulation behaviors and demonstrates superior energy-saving performance in complex driving scenarios.

The purpose of texture measurement is to describe and quantify the texture features of pixels in an image. The accuracy of texture measurement plays a crucial role in determining the effectiveness of texture filtering. However, current texture measurement methods face challenges in achieving accurate texture measurement results, particularly for multi-scale texture measurements. This limitation often leads to unsatisfactory texture filtering results, particularly with image details and high-contrast textures. We find that when moving the texture measurement regions for pixels near texture edges further away from the texture edge and keeping the texture measurement regions for pixels far from texture edges unchanged results in an improved accuracy of texture measurement. Based on this observation, we propose a novel texture measurement approach that employs a circular neighborhood with a variable radius as the texture measurement region for each pixel. Furthermore, we proposed an image terrain map model based on a one-pixel texture edge to obtain optimal parameters for texture measurement regions. This model significantly enhances the accuracy of texture measurement at any scale in an image. The experimental results show that the texture filtering method based on our image terrain map model is significantly better than existing methods in terms of edge-preservation, small-structure preservation, and high-contrast texture filtering. Additionally, we presented some applications of the image terrain map model in other areas of image processing to demonstrate its versatility.

Given that traditional planning methods cannot generate safe and smooth trajectories in complex and unstructured scenarios, one of the crucial challenges is to enable self-driving vehicles to mimic human manipulation behaviors as much as possible. This feature entails the capability to control vehicles safely and smoothly in continuously changing scenarios. Thus, this study proposes a trajectory planning method for humanoid control. First, several professional drivers are invited to maneuver the model vehicle and collect data while driving under obstacle avoidance in outdoor multi-obstacle scenarios. Thereafter, the collected data are processed to establish a dataset representing human manipulation habits. Then, an end-to-end learning approach is used to establish the mapping relationship between the feasible domain space and human manipulation of trajectories. The ResNet compensation attention mechanism (R-CAM) network is used to improve the accuracy of trajectory predictions. Finally, the results of the experimental test analysis with the traditional planning method, two sota methods, and multiple experimental methods show that the trajectory planned by the proposed method closely aligns with human manipulation habits and has superior smoothness. Compared with the baseline method, the root mean square error (RMSE) between the planned trajectory and the real trajectory value in this paper is reduced by 95.5% and 98.8% in the x and y directions, respectively, with higher trajectory precision, and the Var of the angular velocity is decreased by 87.3%, with better smoothness of the planned trajectory.

The process of recovering dense depth maps from sparse depth information is prone to edge blurring. This paper proposes an image-guided depth completion algorithm to address this issue. The method uses the edges of the color image as prior constraints to construct an image topography map as an intermediate representation and performs nonlinear adaptive reconstruction based on the image content to adjust the position and scale of the pixel-weighted neighborhood. This approach avoids incorporating depth information with different distributions when estimating missing values, resulting in a full-resolution dense depth map with sharp edges. We conducted a quantitative comparison with state-of-the-art models on the KITTI and MidAir datasets, demonstrating that our algorithm has better performance and robustness in terms of completion accuracy. We also analyzed the impact of sparsity on the algorithm’s performance and its ability to recover fine structures in dense depth results and demonstrated the reconstruction results for sparse data in real-world scenarios.

This paper introduces a novel multi-sensor fusion SLAM algorithm named FLARE-SLAM, designed for mobile robots operating in complex environments. This algorithm addresses challenges associated with uneven LiDAR measurement signals and their random distribution. First, we enhance the stability of feature extraction by refining the curvature calculation strategy for LiDAR point clouds and incorporating contextual information from the sensor array. Second, we introduce an adaptive residual optimization weight distribution mechanism, grounded in the principle of uniform residual optimization, to boost the algorithm’s adaptability across various environments. Extensive evaluations on the KITTI dataset confirm that FLARE-SLAM constructs a global map with enhanced consistency and accuracy, achieving an absolute trajectory error of 0.53% and an absolute rotation error of 0.19∘/100m. Additionally, we validate the robustness of the algorithm through real-world testing in diverse outdoor and indoor settings.

Most current light detection and ranging (LiDAR)-based simultaneous localization and mapping (SLAM) methods are based on static conditions, but real-world scenarios often violate this prior assumption. To address the existing challenges, this article proposes DO-Removal, an online LiDAR-inertial odometry that removes dynamic objects. Specifically, the method uses ground fitting results as a reference, takes point cloud measurements with significant geometric features as seed points for region growing, and uses clustering results to determine the confidence of dynamic element point cloud segmentation, thereby separating dynamic and static elements. Additionally, this article proposes a multiline LiDAR point cloud feature extraction method that considers context beams simultaneously, enhancing the significance of the extraction results. It also implements a residual optimization function based on distance truncation, distinguishing contributions by confidence, and adaptively weighting features at different distances. Finally, extensive testing was conducted on the KITTI dataset and a self-collected dataset, achieving competitive results with absolute trajectory error and absolute rotation error reduced to 0.51% and 0.19 degrees/100 m, respectively.

Accurately freespace detection is crucial to ensure the safe operation of autonomous vehicles. However, creating multi-scene datasets can be challenging. Mainstream research primarily addresses driving scenes in urban settings while neglecting other types of road environments. This results in a constrained application environment for current freespace detection methods. This paper proposes an adaptive environment scale freespace detection method in 2D image space. The method does not require data labeling and has better environmental adaptability. The core idea is to adaptively map a fixed point cloud scale in 3D space to a pixel scale in 2D space using the camera projection relation to obtain the fine environmental gradient. Then design search rule to label freespace in 2D space. Experiments on two public datasets, urban and field, achieved F1 scores of 92.50% and 89.09%, respectively. In both structured and unstructured environments, the proposed method demonstrated higher accuracy and lower false detection rates compared to state-of-the-art methods.

This article introduces a novel multisensor fusion simultaneous localization and mapping (SLAM) algorithm named FLARE-SLAM, designed for mobile robots operating in complex environments. This algorithm addresses challenges associated with uneven light detection and ranging (LiDAR) measurement signals and their random distribution. First, we enhance the stability of feature extraction by refining the curvature calculation strategy for LiDAR point clouds and incorporating contextual information from the sensor array. Second, we introduce an adaptive residual optimization weight distribution mechanism, grounded in the principle of uniform residual optimization, to boost the algorithm's adaptability across various environments. Extensive evaluations on the KITTI dataset confirm that FLARE-SLAM constructs a global map with enhanced consistency and accuracy, achieving an absolute trajectory error of 0.53% and an absolute rotation error of 0.19 degrees/100 m. Additionally, we validate the robustness of the algorithm through real-world testing in diverse outdoor and indoor settings.

Due to the divergence of accuracy caused by inertial measurement unit (IMU) cumulative error, it is difficult for a single IMU equipment to realize vehicle positioning. Therefore, this paper proposes an IMU pose state estimation algorithm based on modulation long short-term memory-unscented Kalman filter (ML-UKF) algorithm. First, the algorithm improves the memory mode of LSTM network by using Modulation LSTM neural network and establishes IMU state model and observation model. Then, in order to adapt to the application of deep learning algorithm in UKF, an equal spacing sigma sampling method is proposed. Finally, the effect of IMU pose state estimation is verified by experiments. Results show that the root mean square error of the ML-UKF algorithm is decreases by 65.43% relative to the state of the art, further verifying the effectiveness of the proposed algorithm.

Precise and robust extrinsic parameter calibration is fundamental for LiDAR-camera multi-modal sensing applications. However, most existing methods assume that sensors have the same orientation, limiting their effectiveness in feature extraction and feature alignment from different angle of view in multi-angle sensing scenarios. Moreover, the calibration accuracy of existing methods is insufficient for high-performance applications. To address these limitations, we propose a novel automatic extrinsic parameter calibration method utilizing a spherical target. We propose the Curvature Consistency Spherical Detection (CCSD) algorithm for LiDAR point cloud sphere recognition. The CCSD leverages the sphere's structural attributes, enabling robust detection against noise and partial occlusion. To improve camera sphere detection, we present an enhanced ellipse detection technique and compensate the eccentricity error arising from spherical projection based on the principle of perspective transformation. Extensive simulations and real-world experiments demonstrate the proposed method's superiority in accuracy and practicality over state-of-the-art (SOTA) methods.

In this article, we proposed a novel binocular positioning method based on a dynamic optical center imaging model to improve the accuracy of binocular positioning. By analyzing the distribution rules of the optical center at various object distances, we construct a new optical imaging model that is better suited for practical binocular positioning tasks. In addition, we develop a corresponding calibration method to accurately determine the model parameters. The experimental results demonstrate that our binocular positioning method outperforms existing methods in terms of spatial positioning and 3-D reconstruction accuracy. Compared to the binocular positioning method based on the traditional pinhole imaging model, our method achieves an 89.8% enhancement in spatial positioning accuracy and a 96.1% improvement in 3-D reconstruction accuracy for target objects. These results present the effectiveness and superiority of our method in binocular positioning applications.