Ajian Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ajian Liu has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Liang Yuan
USC-SJTU Institute of Cultural and Creative Industry, Shanghai Jiao Tong University, Shanghai, China
4 shared publications - Juan Chen
College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China
2 shared publications - Jun Wang
College of Information Science & Technology, Beijing University of Chemical Technology, Beijing, China
2 shared publications - Lihui Sun
Space Engineering University, China
2 shared publications - Shangzhe Wu
Beijing University of Chemical Technology, China
2 shared publications - Xiaoyu Ni
School of Mechanical Engineering, Hebei University of Architecture, Zhangjiakou, China
2 shared publications
Latest Publications
- 7 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (12) , 18692-18707
UVFusion: A Unified Framework for Voxel-Aligned LiDAR–Camera Fusion in 3-D Object Detection
Accurate three-dimensional (3D) object detection for autonomous driving often relies on the fusion of Light Detection and Ranging (LiDAR) and camera data. However, existing multimodal detectors still suffer from two key limitations: inaccurate cross-modal feature alignment and coarse-grained region-of-interest (RoI) fusion. These limitations are particularly severe for small or distant objects. In addition, most existing methods address feature alignment and RoI fusion separately rather than in a unified framework. This separate treatment makes it difficult to fully exploit their combined benefits for detection performance. To address these problems, we propose UVFusion, a unified framework for voxel-aligned multimodal fusion that integrates image enhancement, cross-modal alignment, and 3D RoI fusion. First, an Adaptive Gated Fusion (AGF) module enhances the image branch with LiDAR-derived geometric priors to generate depth-aware features. Second, a Deformable Cross-Modal Attention (DCMA) module establishes precise correspondences between multi-scale voxel and image features via centroid-guided deformable sampling. Finally, a Grid-level Self-Attention Fusion (GSAF) module performs RoI-wise self-attentive fusion of geometry-enhanced multimodal features, yielding more discriminative 3D representations for proposal refinement. Experiments on the KITTI benchmark show that UVFusion consistently outperforms the Voxel-RCNN baseline, with a 3.06% improvement in 3D average precision (AP) for cars at the hard level, along with significant gains for pedestrians and cyclists. These results demonstrate the effectiveness of the proposed method in detecting small-scale objects in challenging urban scenes.
- 7 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (12) , 18692-18707
UVFusion: A Unified Framework for Voxel-Aligned LiDAR–Camera Fusion in 3-D Object Detection
Accurate three-dimensional (3D) object detection for autonomous driving often relies on the fusion of Light Detection and Ranging (LiDAR) and camera data. However, existing multimodal detectors still suffer from two key limitations: inaccurate cross-modal feature alignment and coarse-grained region-of-interest (RoI) fusion. These limitations are particularly severe for small or distant objects. In addition, most existing methods address feature alignment and RoI fusion separately rather than in a unified framework. This separate treatment makes it difficult to fully exploit their combined benefits for detection performance. To address these problems, we propose UVFusion, a unified framework for voxel-aligned multimodal fusion that integrates image enhancement, cross-modal alignment, and 3D RoI fusion. First, an Adaptive Gated Fusion (AGF) module enhances the image branch with LiDAR-derived geometric priors to generate depth-aware features. Second, a Deformable Cross-Modal Attention (DCMA) module establishes precise correspondences between multi-scale voxel and image features via centroid-guided deformable sampling. Finally, a Grid-level Self-Attention Fusion (GSAF) module performs RoI-wise self-attentive fusion of geometry-enhanced multimodal features, yielding more discriminative 3D representations for proposal refinement. Experiments on the KITTI benchmark show that UVFusion consistently outperforms the Voxel-RCNN baseline, with a 3.06% improvement in 3D average precision (AP) for cars at the hard level, along with significant gains for pedestrians and cyclists. These results demonstrate the effectiveness of the proposed method in detecting small-scale objects in challenging urban scenes.
- 23 May 2025
- Published by Elsevier in Neurocomputing
- Vol. 645
Talent3D: Optimizing geometry and enhancing appearance for high-quality Text-to-3D shape generation
- 23 May 2025
- Published by Elsevier in Neurocomputing
- Vol. 645
Talent3D: Optimizing geometry and enhancing appearance for high-quality Text-to-3D shape generation
- 9 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-13
CSA-RCNN: Cascaded Self-Attention Networks for High-Quality 3-D Object Detection From LiDAR Point Clouds
LiDAR-based 3-D object detection leverages the precise spatial information provided by point clouds to enhance the understanding of 3-D environments. This approach has garnered significant attention from both industry and academia, particularly in fields such as autonomous driving and robotics. However, how to improve the detection accuracy of long-distance objects remains a critical challenge for existing two-stage methods. This difficulty is primarily attributed to the sparsity and uneven distribution of point clouds, which lead to inconsistent quality in proposals for distant targets. To tackle the challenges, this paper proposes a novel and effective 3-D point cloudy detection network based on cascaded self-attention, named CSA-RCNN, to achieve higher quality 3-D object detection in traffic scenes. First, to enhance the quality of proposals for long-range objects, we design a cascaded self-attention module (CSM) that utilizes a multi-head self-attention (MHSA) mechanism across multiple independent cascaded sub-networks to aggregate proposal features at different stages. This approach improves the accuracy of iterative proposal refinement by strengthening the feature modeling across different stages. Second, to enhance the correlation between different representations of the point cloud, we design a transformer-based feature fusion module that fully integrates these multi-source features into richer point-wise features. Finally, to remove unnecessary background information from the 3-D scene, we introduce a Semantic-guided Farthest Point Sampling (S-FPS) strategy that helps preserve essential foreground points during the downsampling process. Extensive experiments were conducted on the highly competitive KITTI and Waymo datasets, which validated the effectiveness of the proposed method. Notably, CSA-RCNN achieves a +1.01% improvement in Average Precision (AP) for the car class at the difficult level, compared to the PV-RCNN on the KITTI validation dataset.
- 9 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-13
CSA-RCNN: Cascaded Self-Attention Networks for High-Quality 3-D Object Detection From LiDAR Point Clouds
LiDAR-based 3-D object detection leverages the precise spatial information provided by point clouds to enhance the understanding of 3-D environments. This approach has garnered significant attention from both industry and academia, particularly in fields such as autonomous driving and robotics. However, how to improve the detection accuracy of long-distance objects remains a critical challenge for existing two-stage methods. This difficulty is primarily attributed to the sparsity and uneven distribution of point clouds, which lead to inconsistent quality in proposals for distant targets. To tackle the challenges, this paper proposes a novel and effective 3-D point cloudy detection network based on cascaded self-attention, named CSA-RCNN, to achieve higher quality 3-D object detection in traffic scenes. First, to enhance the quality of proposals for long-range objects, we design a cascaded self-attention module (CSM) that utilizes a multi-head self-attention (MHSA) mechanism across multiple independent cascaded sub-networks to aggregate proposal features at different stages. This approach improves the accuracy of iterative proposal refinement by strengthening the feature modeling across different stages. Second, to enhance the correlation between different representations of the point cloud, we design a transformer-based feature fusion module that fully integrates these multi-source features into richer point-wise features. Finally, to remove unnecessary background information from the 3-D scene, we introduce a Semantic-guided Farthest Point Sampling (S-FPS) strategy that helps preserve essential foreground points during the downsampling process. Extensive experiments were conducted on the highly competitive KITTI and Waymo datasets, which validated the effectiveness of the proposed method. Notably, CSA-RCNN achieves a +1.01% improvement in Average Precision (AP) for the car class at the difficult level, compared to the PV-RCNN on the KITTI validation dataset.
- Open Access
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Three-Dimensional Object Segmentation Method based on YOLO, SAM, and NeRF
The neural radiance field (NeRF) representation has shown promising results in capturing 3D scenes. However, when it comes to editing or moving specific objects within virtual environments, it is essential to segment individual objects from the neural radiance field. Existing methods primarily rely on feature computation and clustering for 3D segmentation, but they often suffer from poor segmentation quality and low accuracy. Taking advantage of recent advancements in 2D image segmentation, we propose a 3D segmentation method that generates single-object segmentation from multiple viewpoint images. First, we compute camera poses using feature points from the original images. Then, with user guidance, we use the You Only Look Once (YOLO) network model to roughly locate the objects and utilize the resulting information as hints. Subsequently, we employ the Segment Anything Model (SAM) for fine-grained segmentation. The processed dataset is then rendered into a 3D segmented object using the neural radiance field, producing high-quality rendered images. By training on specific objects, we can segment object categories that the publicly available YOLO model cannot recognize, thereby enhancing generalization. Through experiments, we analyze the real-time performance of each module and qualitatively and quantitatively evaluate the effectiveness of our method. Compared to the N3F and ISRF methods, our approach achieves an average 17% improvement in Intersection over Union (IoU) and an average 9% improvement in Pixel Accuracy (Acc). We validate the performance of our method on the LLFF, mip-Nerf360 datasets, demonstrating its generalizability and effectiveness.
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessThree-Dimensional Object Segmentation Method based on YOLO, SAM, and NeRF
The neural radiance field (NeRF) representation has shown promising results in capturing 3D scenes. However, when it comes to editing or moving specific objects within virtual environments, it is essential to segment individual objects from the neural radiance field. Existing methods primarily rely on feature computation and clustering for 3D segmentation, but they often suffer from poor segmentation quality and low accuracy. Taking advantage of recent advancements in 2D image segmentation, we propose a 3D segmentation method that generates single-object segmentation from multiple viewpoint images. First, we compute camera poses using feature points from the original images. Then, with user guidance, we use the You Only Look Once (YOLO) network model to roughly locate the objects and utilize the resulting information as hints. Subsequently, we employ the Segment Anything Model (SAM) for fine-grained segmentation. The processed dataset is then rendered into a 3D segmented object using the neural radiance field, producing high-quality rendered images. By training on specific objects, we can segment object categories that the publicly available YOLO model cannot recognize, thereby enhancing generalization. Through experiments, we analyze the real-time performance of each module and qualitatively and quantitatively evaluate the effectiveness of our method. Compared to the N3F and ISRF methods, our approach achieves an average 17% improvement in Intersection over Union (IoU) and an average 9% improvement in Pixel Accuracy (Acc). We validate the performance of our method on the LLFF, mip-Nerf360 datasets, demonstrating its generalizability and effectiveness.
- Open Access
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-7
An Object Detection Algorithm Based on Multiple Attention Mechanisms
In the field of autonomous driving, a novel approach is proposed for detecting and recognizing surrounding vehicles targets. The proposed method combines multiple attention mechanisms within a 3D object detection algorithm, drawing inspiration from the PointPillars network. Firstly, an additional local attention mechanism layer is incorporated into the point cloud pillar feature extraction module. This augmentation allows for a more comprehensive fusion of surrounding point cloud feature during the point cloud feature extraction process. Secondly, a channel attention mechanism module is introduced to the 2D image backbone network. The resulting features are subsequently inputted into the SSD algorithm for object detection. The improved algorithm's effectiveness has been demonstrated through an experimental evaluation on the popular KITTI dataset. The results show a significant improvement in average mean accuracy compared to the original algorithm, indicating that the attention mechanisms successfully enhance the detection and recognition performance of surrounding vehicles in autonomous driving scenarios.
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessAn Object Detection Algorithm Based on Multiple Attention Mechanisms
In the field of autonomous driving, a novel approach is proposed for detecting and recognizing surrounding vehicles targets. The proposed method combines multiple attention mechanisms within a 3D object detection algorithm, drawing inspiration from the PointPillars network. Firstly, an additional local attention mechanism layer is incorporated into the point cloud pillar feature extraction module. This augmentation allows for a more comprehensive fusion of surrounding point cloud feature during the point cloud feature extraction process. Secondly, a channel attention mechanism module is introduced to the 2D image backbone network. The resulting features are subsequently inputted into the SSD algorithm for object detection. The improved algorithm's effectiveness has been demonstrated through an experimental evaluation on the popular KITTI dataset. The results show a significant improvement in average mean accuracy compared to the original algorithm, indicating that the attention mechanisms successfully enhance the detection and recognition performance of surrounding vehicles in autonomous driving scenarios.
- Open Access
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Semantic SLAM with Geometric Constraints for Object Tracking in Dynamic Environments
Visual simultaneous localization and mapping (VSLAM) is extensively utilized in the field of unmanned driving. Classical VSLAM typically postulates that the surroundings are stationary, and dynamic objects can seriously affect the accuracy of the system. However, detecting and tracking dynamic objects is beneficial for improving the autonomy and intelligence of robotics. In order to solve the problem of VSLAM distinguishing dynamic objects in dynamic scenes, a semantic VSLAM system based on the VDO-SLAM framework is proposed. The approach integrates semantic segmentation into the visual odometry and combines scene flow and geometric constraints to effectively reduce the negative impact of dynamic objects on system performance. Specifically, a transformer-based neural network architecture for optical flow estimation was employed, which effectively enhances the segmentation precision of foreground and background regions. And a geometric constraint was introduced to classify dynamic objects, which can reduction of erroneous segmentation of dynamic objects with effect. The data analysis results on the KITTI dataset demonstrate that our proposed algorithms can effectively enhance precision and stability of VSLAM system, reduce the tracking error of VSLAM in dynamic environments, and demonstrate good potential and application value.
- 25 August 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessSemantic SLAM with Geometric Constraints for Object Tracking in Dynamic Environments
Visual simultaneous localization and mapping (VSLAM) is extensively utilized in the field of unmanned driving. Classical VSLAM typically postulates that the surroundings are stationary, and dynamic objects can seriously affect the accuracy of the system. However, detecting and tracking dynamic objects is beneficial for improving the autonomy and intelligence of robotics. In order to solve the problem of VSLAM distinguishing dynamic objects in dynamic scenes, a semantic VSLAM system based on the VDO-SLAM framework is proposed. The approach integrates semantic segmentation into the visual odometry and combines scene flow and geometric constraints to effectively reduce the negative impact of dynamic objects on system performance. Specifically, a transformer-based neural network architecture for optical flow estimation was employed, which effectively enhances the segmentation precision of foreground and background regions. And a geometric constraint was introduced to classify dynamic objects, which can reduction of erroneous segmentation of dynamic objects with effect. The data analysis results on the KITTI dataset demonstrate that our proposed algorithms can effectively enhance precision and stability of VSLAM system, reduce the tracking error of VSLAM in dynamic environments, and demonstrate good potential and application value.