Ziming Zhu

Mr. Ziming Zhu

orcid

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
27 November 2023 in Sensors
Verified
Journal Article
Open access
PDF file
SSShi Shi
Shi Shi
ZZZiming Zhu
Ziming Zhu
KZKezhi Zhang
Kezhi Zhang
HCHuanlei Chen
Huanlei Chen
ZYZhuoping Yu
Zhuoping Yu
YZYu Zhu
Yu Zhu
4D millimeter wave (mmWave) imaging radar is a new type of vehicle sensor technology that is critical to autonomous driving systems due to its lower cost and robustness in complex weather. However, the sparseness and noise of point clouds are still the main problems restricting the practical application of 4D imaging radar. In this paper, we introduce SMIFormer, a multi-view feature fusion network framework based on 4D radar single-modal input. SMIFormer decouples the 3D point cloud scene into 3 independent but interrelated perspectives, including bird’s-eye view (BEV), front view (FV), and side view (SV), thereby better modeling the entire 3D scene and overcoming the shortcomings of insufficient feature representation capabilities under single-view built from extremely sparse point clouds. For multi-view features, we proposed multi-view feature interaction (MVI) to exploit the inner relationship between different views by integrating features from intra-view interaction and cross-view interaction. We evaluated the proposed SMIFormer on the View-of-Delft (VoD) dataset. The mAP of our method reached 48.77 and 71.13 in the fully annotated area and the driving corridor area, respectively. This shows that 4D radar has great development potential in the field of 3D object detection.

4 November 2024 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
ZZZiming Zhu
Ziming Zhu
YZYu Zhu
Yu Zhu
KZKezhi Zhang
Kezhi Zhang
HLHangyu Li
Hangyu Li
XLXiaofeng Ling
Xiaofeng Ling
The effective representation and feature extraction of 3D scenes from sparse and unstructured point clouds pose a significant challenge in 3D object detection. In this paper, we propose TransMRE, a network that enables fully sparse multiple observation plane feature fusion using LiDAR point clouds as single-modal input. TransMRE achieves this by sparsely factorizing a 3D voxel scene into three separate observation planes: X-Y plane, X-Z plane, and Y-Z plane. Additionally, we propose Observation Plane Sparse Fusion and Interaction to explore the internal relationship between different observation planes. The transformers mechanism is employed to realize feature attention within single observation plane and feature attention across multiple observation planes. This recursive application of attention is done during multiple observation plane projection feature aggregation to effectively model the entire 3D scene. This approach addresses the limitation of insufficient feature representation ability under a single bird’s-eye view constructed by extremely sparse point clouds. Furthermore, TransMRE maintains the full sparsity property of the entire network, eliminating the need to convert sparse feature maps into dense feature maps. As a result, it can be effectively applied to LiDAR point cloud data with large scanning ranges, such as Argoverse 2, while ensuring low computational complexity. Extensive experiments were conducted to evaluate the effectiveness of TransMRE, achieving 64.9 mAP and 70.4 NDS on the nuScenes detection benchmark, and 32.3 mAP on the Argoverse 2 detection benchmark. These results demonstrate that our method outperforms state-of-the-art methods.

1 January 2026 in IEEE Sensors Journal
Verified
Journal Article
Closed access
JCJiahao Chen
Jiahao Chen
HCHuanlei Chen
Huanlei Chen
ZZZiming Zhu
Ziming Zhu
ZSZheng Shen
Zheng Shen
XLXiaofeng Ling
Xiaofeng Ling
YZYu Zhu
Yu Zhu
The effective representation and feature extraction from sparse point clouds of 4D millimeter-wave(4D-MMW) radar pose a significant challenge in 3D object detection. This paper proposes DDCFusion, a novel radar-camera fusion network that advances measurement precision by dynamically compensating for depth errors in sparse radar data. DDCFusion achieves this by exploiting the physical properties of 4D-MMW radar to improve measurement reliability and reduce depth uncertainty, which enhances depth measurement confidence in the view transform by integrating RCS-derived reflectivity metrics. The occupancy-weighted radar branch prioritizes image regions with high-confidence radar returns, minimizing measurement noise in view transform operation. Furthermore, DDCFusion optimizes spatial measurement consistency in Bird’s-Eye-View (BEV) space by modeling cross-sensor dependencies through the Global Feature Slice Coordinate Attention (GFSCA) fusion module. Experimental validation on the VoD and TJ4DRadSet datasets demonstrates superior measurement accuracy, achieving 51.08% mAP on VoD and 34.61% mAP on TJ4DRadSet—outperforming existing methods in depth error reduction and robustness to sparsity. Ablation studies verify the measurement-centric design: RCS-guided diffusion improves small-object detection (e.g., pedestrians), while DBSCAN-based clustering refines large-object localization (e.g., vehicles). The network demonstrates significant improvements in depth accuracy and robustness to sparse inputs while maintaining competitive inference latency with 138ms.

28 October 2021 in Sensors
Verified
Journal Article
Open access
PDF file
GXGuansheng Xing
Guansheng Xing
ZZZiming Zhu
Ziming Zhu
Lane and road marker segmentation is crucial in autonomous driving, and many related methods have been proposed in this field. However, most of them are based on single-frame prediction, which causes unstable results between frames. Some semantic multi-frame segmentation methods produce error accumulation and are not fast enough. Therefore, we propose a deep learning algorithm that takes into account the continuity information of adjacent image frames, including image sequence processing and an end-to-end trainable multi-input single-output network to jointly process the segmentation of lanes and road markers. In order to emphasize the location of the target with high probability in the adjacent frames and to refine the segmentation result of the current frame, we explicitly consider the time consistency between frames, expand the segmentation region of the previous frame, and use the optical flow of the adjacent frames to reverse the past prediction, then use it as an additional input of the network in training and reasoning, thereby improving the network’s attention to the target area of the past frame. We segmented lanes and road markers on the Baidu Apolloscape lanemark segmentation dataset and CULane dataset, and present benchmarks for different networks. The experimental results show that this method accelerates the segmentation speed of video lanes and road markers by 2.5 times, increases accuracy by 1.4%, and reduces temporal consistency by only 2.2% at most.


25 December 2025 in Research Square
Verified
Preprint
Open access
PDF file
YZYu Zhu
Yu Zhu
SWShengze Wang
Shengze Wang
YGYufeng Gu
Yufeng Gu
ZZZiming Zhu
Ziming Zhu
NWNan Wang
Nan Wang
With the proliferation of deepfake techniques, face forgery detection has emerged as a critical area of research to mitigate the risks associated with facial manipulation. Recent studies have highlighted the superiority of frequency information over color-space input, especially in high-compression scenarios. However, most frequency-enhanced methods employ a two-stream network architecture, assigning separate encoders for RGB and frequency domains, and rely solely on vanilla binary cross-entropy loss, limiting their generalization ability.To address these limitations, we propose a novel framework, Frequency-enhanced Contrastive Learning (FCL), which trains the model in a supervised manner using contrastive loss. FCL treats two views of different modalities generated from the same image as positive pairs and samples with opposite labels as negative pairs. This approach pulls features from the RGB and high-frequency domains closer while pushing features of pristine and forgery faces apart. Additionally, we introduce a Shallow Feature Supplement (SFS) module to complement local information from low-level shallow feature embeddings into high-level feature maps, and a Dual Modal Fusion (DMF) module to adaptively aggregate information from both domains.Extensive experiments on seven datasets demonstrate the superior generalization of our method compared to state-of-the-art competitors. Notably, FCL achieves a remarkable accuracy of 92.29% and an AUC of 94.25% on the heavily compressed FaceForensics++ dataset, showcasing its robustness against compression. These results underscore the potential of FCL as a powerful tool for face forgery detection in real-world scenarios.The source code ofour proposed algorithm are available at https://github.com/Y30230924Wang/Frequency-Enhanced-Contrastive-Learning.