Article
Robotics
Shihao Feng, Pengpeng Liang, Jin Gao, Erkang Cheng
Summary: This letter presents a point cloud-based 3D object tracking method using a multi-correlation Siamese Transformer network. Each stage of the network performs feature correlation based on sparse pillars, enabling effective learning of the correlation between the template and search branches while preserving individual characteristics. The proposed algorithm achieves promising performance compared to state-of-the-art methods on popular datasets.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2023)
Article
Computer Science, Artificial Intelligence
Ruixu Wu, Xianbin Wen, Liming Yuan, Haixia Xu
Summary: In this paper, a new object tracking method (DASFTOT) is proposed, which integrates backbone network, transformer mechanism, and bounding prediction box. The method shows high effectiveness in fusing local and global features, calculating the correlation between templates and search regions, and achieving comparable results to state-of-the-art methods.
KNOWLEDGE-BASED SYSTEMS
(2022)
Article
Computer Science, Information Systems
Lin Zhang, Hua Meng, Yunbing Yan, Xiaowei Xu
Summary: TGPP algorithm is an improved method of PointPillars algorithm, which utilizes Transformer and multi-head attention mechanism to extract global contextual features and local structure features of point cloud, thus improving the performance of object detection. It achieved an average accuracy improvement of 2.64% in the KITTI test set.
Article
Engineering, Electrical & Electronic
Zheng Fang, Sifan Zhou, Yubo Cui, Sebastian Scherer
Summary: This paper introduces a 3D tracking method called 3D-SiamRPN Network, which tracks a single target object using raw 3D point cloud data. Experimental results show its competitive performance in both Success and Precision, as well as real-time running capabilities.
IEEE SENSORS JOURNAL
(2021)
Article
Computer Science, Artificial Intelligence
Xu Wang, Yuqiao Zeng, Yi Jin, Yigang Cen, Baifu Liu, Shaohua Wan
Summary: This paper proposes a novel architecture called Relation-Shape Transformer Network (RS-TNet) for 3D point cloud representation. By integrating global multi-head self-attention and local relation-feature extraction module, RS-TNet achieves coarse-to-fine grained semantic information coverage, introduces spatial relation of points by learning underlying shapes, and generates features with more shape awareness and robustness.
Article
Computer Science, Artificial Intelligence
Xin Chen, Bin Yan, Jiawen Zhu, Huchuan Lu, Xiang Ruan, Dong Wang
Summary: A novel attention-based feature fusion network inspired by the transformer is proposed for template matching in the tracking field. The experiments demonstrate that the proposed method achieves promising results on seven popular benchmark datasets.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
(2023)
Article
Chemistry, Multidisciplinary
Jiaqi Xi, Jin Yang, Xiaodong Chen, Yi Wang, Huaiyu Cai
Summary: This study proposes a Siamese tracker equipped with a Double Branch Attention (DBA) block to suppress the influence of background distractors on target representation in object tracking. By concatenating channels from multiple layers and applying channel attention and contextual relevance, this method enhances the discriminative representation of the target. Experimental results show that this approach significantly improves tracking performance.
APPLIED SCIENCES-BASEL
(2022)
Article
Computer Science, Artificial Intelligence
Yuxuan Wang, Liping Yan, Zihang Feng, Yuanqing Xia, Bo Xiao
Summary: In the field of single object tracking, cross-correlation operation is crucial for Siamese-based trackers. However, the traditional cross-correlation method tends to have insufficient utilization or even loss of feature information due to the local linear matching of the search region using target features. To address this issue, we propose a novel matching operator inspired by Transformer, which utilizes multi-head attention and a designed modulation module. Additionally, our tracker incorporates a multi-scale encoder/decoder strategy for improved tracking accuracy. The proposed tracker, named VTTR, achieves excellent performance on multiple benchmarks while maintaining fast speed.
IMAGE AND VISION COMPUTING
(2023)
Article
Automation & Control Systems
Kai Huang, Jun Chu, Lu Leng, Xingbo Dong
Summary: This study proposes a Transformer-based Siamese tracking architecture called TATrack, which integrated with deformable attention to focus on the relevant information about the target, thereby improving the performance of object tracking.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2024)
Article
Engineering, Civil
Kanglin Ning, Yanfei Liu, Yanzhao Su, Ke Jiang
Summary: This paper proposes a 3D object detection framework based on point-voxel and bird's-eye-view representation aggregation network. It integrates embedded Fourier features and features extracted by a convolution backbone, and dynamically fuses spatial-level and semantic-level features to improve detection accuracy.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2023)
Article
Computer Science, Hardware & Architecture
Xiaolou Sun, Qi Wang, Fei Xie, Zhibin Quan, Wei Wang, Hao Wang, Yuncong Yao, Wankou Yang, Satoshi Suzuki
Summary: This paper introduces a new tracking system framework SiamTrans, which achieves a balance between speed and accuracy through Siamese Transformer Network, and enhances the robustness of the tracking system by utilizing Tracking Drift Suppression Strategy. Experimental results demonstrate competitive performance of SiamTrans across multiple benchmarks, and real-time requirements can be met on embedded devices as well.
JOURNAL OF SYSTEMS ARCHITECTURE
(2022)
Article
Engineering, Civil
Xin Cheng, Jingmei Zhou, Peiyuan Liu, Xiangmo Zhao, Hongfei Wang
Summary: This paper proposes a 3D vehicle object tracking algorithm based on bounding box similarity measurement, which effectively extracts features from a discrete point cloud for vehicle tracking. The algorithm utilizes a 3D Kalman filter and greedy matching to achieve 3D vehicle object tracking.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2023)
Article
Environmental Sciences
Chunhui Zhao, Hongjiao Liu, Nan Su, Congan Xu, Yiming Yan, Shou Feng
Summary: This study proposes a Transformer-based multimodality information transfer network (TMTNet) to improve hyperspectral object tracking by efficiently transferring multimodality data information composed of RGB and hyperspectral data. Two subnetworks are constructed to transfer multimodality fusion information and robust RGB visual information, respectively. The proposed TMTNet tracker outperforms advanced trackers, demonstrating its effectiveness.
Article
Computer Science, Information Systems
Weifan Xu, Jin Jin, Fenglei Xu, Ze Li, Chongben Tao
Summary: In the field of autonomous driving, precise spatial positioning and 3D object detection are crucial. Traditional detection models for RGB images struggle with the disorder in LiDAR point clouds. We propose Frustumformer, a novel framework that leverages the inherent order of LiDAR point clouds to enhance representation. It effectively models long-range dependencies and outperforms existing methods in experiments on the KITTI dataset.
Article
Engineering, Electrical & Electronic
Xiaoyu Tian, Ming Yang, Qian Yu, Junhai Yong, Dong Xu
Summary: This paper proposes MedoidsFormer, a novel transformer-based backbone that is specifically designed for LiDAR-based 3D object detection. With the introduction of Medoids Attention, an innovative self-attention mechanism, the model can exploit interactions within surrounding regions, reduce computation and memory costs, and obtain discriminative context information. Extensive experiments demonstrate consistent improvement over existing 3D object detectors and state-of-the-art performance on the large-scale Waymo Open Dataset.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2023)
Article
Chemistry, Analytical
Yubo Cui, Zheng Fang, Sifan Zhou
Article
Engineering, Electrical & Electronic
Zheng Fang, Sifan Zhou, Yubo Cui, Sebastian Scherer
Summary: This paper introduces a 3D tracking method called 3D-SiamRPN Network, which tracks a single target object using raw 3D point cloud data. Experimental results show its competitive performance in both Success and Precision, as well as real-time running capabilities.
IEEE SENSORS JOURNAL
(2021)
Article
Robotics
Yubo Cui, Jiayao Shan, Zuoxu Gu, Zhiheng Li, Zheng Fang
Summary: This paper proposes a sparse-to-dense and transformer-based framework for 3D single object tracking. By transforming sparse points into pillar structures and compressing them into 2D features, a dense representation is obtained. The framework utilizes attention-based computation for global similarity and multi-scale feature compensation. The object tracking is achieved through a two-stage decoder.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2022)
Article
Engineering, Electrical & Electronic
Miao Sun, Yingjie Cao, Jian Qian, Jie Li, Sifan Zhou, Ziyu Zhao, Yifan Wu, Tao Xia, Yajie Qin, Lei Qiu, Shunli Ma, Patrick Yin Chiang, Shenglong Zhuo
Summary: This paper presents a heterogeneous AI-accelerator SoC specifically designed for depth image completion computation. Three key innovations are introduced to enhance the SoC's performance, including a fully-filled dataflow management engine for preprocessing the RGB+Depth input, a hardware-tiling co-processor for improving the efficiency of the CNN accelerator, and the incorporation of a RISC-V core to better execute vector computations. The implemented SoC achieves 2TOPs/W energy efficiency and 34fps throughput under VGA-resolution output for real-time LiDAR systems.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II-EXPRESS BRIEFS
(2023)
Article
Chemistry, Physical
Sifan Zhou, Chunming Yang, Li Guo, Razium Ali Soomro, Maomao Niu, Zhixiong Yang, Rui Du, Danjun Wang, Feng Fu, Bin Xu
Summary: The construction of a FeS2/S-ZnSnO3 heterostructure was carried out to achieve efficient photocatalytic hydrogen evolution reaction (HER) activity. The band structure of ZnSnO3 was regulated by sulfur doping, and FeS2 nanoparticles were coupled to improve optical absorption and carrier separation/transfer in the composite. The optimized heterostructure (8.7%FeS2@S15%-ZSO) showed a HER performance of 2225 μmol g-1 h-1, which was significantly higher than ZSO, S15%-ZSO, and FeS2. DFT calculations confirmed that S doping regulated the electronic structure of S-ZSO, while the coupling of FeS2 constructed a S-Scheme heterostructure, leading to improved HER performance.
APPLIED SURFACE SCIENCE
(2023)