Tengfei Lu

Tengfei Lu

Auto-generated profile, based on publicly available information

Tengfei Lu

School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

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

Representative Publications

View All Publications
20 September 2023 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
TLTengfei Lu
Tengfei Lu
ZWZhong-Li Wang
Zhong-Li Wang
YSYan Shen
Yan Shen
XSXiaotao Shao
Xiaotao Shao
YTYonglin Tang
Yonglin Tang
Catenary is one of the most crucial parts of electrified railway system. How to detect the part defects of catenary in time and keep it in a stable and safe operation conditions are the main tasks for maintainance. The existing data-driven vision-based defect detection methods will face a big challenge that, there are many catenary parts and each part has several types of defects, but there are few defect sample images for each type of defect, which limits them in real applications. To alleviate this problem, a semantic label-enhanced variational autoencoder (VAE) method for catenary part defect detection, termed Defect VAE (DefVAE), is presented. The proposed method is based on a label-enhanced VAE network that determines the distribution boundary in latent feature space of defect sample for additional defect generation. The addition of semantic label information to the VAE improves the inter-class distance in latent space, clarifying the boundary and further boosting the capabilities of the defect detection method, which is demonstrated in our experiments. Additionally, the defect type is classified by combining the outputs of the classifier confidence and the pixel-level reconstruction error which is based on a sliding label mode of the variational autoencoder. Extensive experiments with the open benchmark dataset MVTec and the catenary dataset collected by ourselves demonstrate that the presented DefVAE outperforms the baseline methods across the majority of indicators.

29 May 2025 in IEEE Robotics and Automation Letters
Verified
Journal Article
Closed access
TLTengfei Lu
Tengfei Lu
ZWZhong-Li Wang
Zhong-Li Wang
XFXiaoyang Fan
Xiaoyang Fan
YSYan Shen
Yan Shen
XSXiaotao Shao
Xiaotao Shao
Animals in nature exhibit remarkable spatial cognition abilities, enabling them to achieve long-distance autonomous navigation efficiently in unknown environments. Neurobiologically inspired SLAM methods primarily aim to mimic this capability of spatial cognition and navigation. However, existing works have only been able to perform rough velocity estimation with scanline intensity profiles, which may lead to a severely distorted cognitive map, particularly in large scale environment. This work presents a novel neurobiologically inspired Visual-Inertial SLAM method, called Hybrid-NueroSLAM, that addresses two key issues: precisely and robust visual-inertial fusion-based motion estimation as input for cognition map, and the map consistency for large scale environment. Based on high-precision Visual-Inertial odometry, a precise and robust visual memory mechanism, and a neural cell network based on our proposed improved Bayesian attractor for spatial cognition, it constructs a semi-metric topological experience map. The experimental results with benchmark datasets KITTI and Complex Urban, demonstrate that the proposed method effectively mitigates map jitters and maintains high accuracy and robustness, even on a 11 km urban dataset. Source code: https://github.com/iRobotVision/Hybird-NeuroSLAM

18 January 2024 in IEEE Sensors Journal
Verified
Journal Article
Closed access
DZDonglin Zhu
Donglin Zhu
ZWZhong-Li Wang
Zhong-Li Wang
TLTengfei Lu
Tengfei Lu
XJXin Jiang
Xin Jiang
In a dynamic environment, semantic information can assist the SLAM system in eliminating dynamic point interference. However, most three-dimensional semantic segmentation methods are computationally expensive, which also have low segmentation accuracy for both distant and small objects. We propose a PMF-SLAM method to fully exploit the interaction between 3D semantic segmentation and SLAM and achieve efficient scene perception. The PMF-SLAM system includes three parts: MSF-SegNet model, Interactive SLAM module, Pose-Guiding segmentation module. To improve the accuracy of distant and small objects, MSF-SegNet merges point-wise global features and voxel-wise local features from two branches by a designed symmetrical sparse convolution structure. In the Interactive SLAM module, the coarse-to-fine registration method based on semantic information completes the estimation of pose. To implement the interaction between Segmentation and SLAM, the Pose-Guiding segmentation module was built to assist the segmentation thread in improving inference efficiency and ensuring segmentation consistency over time. Extensive experiments including both local experiment and nuScenes dataset test have been conducted to validate the performance of the proposed method. Our method achieves better accuracy than multiple segmentation algorithms, significantly improving the segmentation performance of distant and small objects. And the trajectory estimation accuracy is better than multiple SLAM algorithms. Code is available at https://github.com/haroldgt/MSF-SegNet.

23 October 2025 in Measurement
Verified
Journal Article
Closed access
DZDonglin Zhu
Donglin Zhu
ZWZhong-Li Wang
Zhong-Li Wang
MCMiao Chen
Miao Chen
TLTengfei Lu
Tengfei Lu
LGLE Gao
LE Gao
YSYan Shen
Yan Shen

10 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
ZLZhongmou Liao
Zhongmou Liao
ZWZhongli Wang
Zhongli Wang
TLTengfei Lu
Tengfei Lu
Event stereo preserves fine temporal information, yet correspondence remains difficult whenmeasurements are sparse, noisy, and unevenly distributed. We present a training-free pipeline that combines pixel-wise leaky integrate-and-fire filtering, adaptive-window robust line fitting, temporal trend smoothing, and offline density-guided interpolation. Its evidence is reassessed with unique ground-truth frame-pixel units, explicit coverage, a 15 ms association tolerance, and no error-dependent rejection. Indoor Flying 1-3 are treated as development data. Indoor Flying 4 is frozen before testing, and an ICCV 2025 EMatch checkpoint provides a contemporary learned baseline under the same evaluator.On the held-out interval, the proposed method obtains 4.438 px native-support mean absolute error (MAE) at 14.35% support. EMatch reaches 0.814 px at 91.57% support. Its advantage remains on the 661,508 units shared by both methods, where their MAEs are 0.687 and 4.432 px. On Indoor Flying 1, the proposed output has lower common-support error than two early local implementations, but that intersection covers only 0.18% of valid units. Registered component reruns show that the adaptivethreshold and window mainly alter support. Trend smoothing lowers common-support MAE by 0.099 px, while densification increases support by a factor of 5.25 without changing predictions already present. Five independent CPU runs give a median full-pipeline time of 490.20 s for 70.29 s of data and 616.0 MiB median peak process-tree memory. The method offers explicit, weight-free inference, but the frozen comparison does not support state-of-the-art accuracy or cross-dataset generalisation.

28 August 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
ZLZhongmou Liao
Zhongmou Liao
ZWZhongli Wang
Zhongli Wang
TLTengfei Lu
Tengfei Lu
Event cameras offer a novel paradigm for stereo vision through their high temporal resolution and high dynamic range, yet the intrinsic asynchrony, sparsity, and high-frequency noise of event data pose significant challenges for robust correspondence estimation. While deep learning-based methods achieve high accuracy, they incur substantial computational overhead; traditional heuristic approaches, in turn, often lack robustness in dynamic scenes. This paper proposes a training-free binocular stereo matching frame-work built on three complementary modules. An adaptive LIF neuron filter with a logarithmic firing threshold suppresses noise while preserving scene structure by modulating sensitivity according to local event density. A robust spatiotemporal plane fitting module uses Iteratively Reweighted Least Squares (IRLS) with a Huber loss and adaptive temporal windows for outlier-resilient disparity estimation. A density-guided interpolation strategy based on spatiotemporal window occupancy then significantly en- hances the spatial completeness of the disparity maps. Evaluated on the Multi Vehicle Stereo Event Camera (MVSEC) dataset,the framework achieves a 49.6% reduction in Mean Absolute Error (MAE) compared with early bio-inspired baselines (Rogister et al., 2012; Firouzi & Conradt, 2016). Under the common 33 ms event pseudo-frame construction and evaluation pipeline used on Indoor Flying 1, the measured results also provide a protocol-specific reference alongside our reproduced RAFT-Stereo baseline. The current Python implementation requires no GPU, has a peak memory footprint of 51.3 MB, and processes the tested sequence at approximately 0.15× real time. Here, training-free denotes the absence of learnable weights and offline model training ratherthan the absence of parameter selection.