ZWZeli Wang

Zeli Wang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zeli Wang has a total of 33 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Dynamic Collision Risk Warning in Construction Sites for Nonlinear Movements: A Visual–Spatial Fusion Multimodal Prediction Method

    by Zeli Wang, Jue Li, Xuzhong Yan, Heng Li, Jia Li

    Collisions between vehicles and workers at construction sites frequently occur at blind corners, where trajectories change abruptly, making it one of the major causes of severe safety incidents. Existing trajectory prediction methods struggle to accurately predict vehicle turning behavior, whereas distance-based collision warning systems frequently generate false alarms. These issues limit their effectiveness in safety management. To address these challenges, this study develops a visual-spatial fusion multimodal trajectory prediction method that can sense turning intent in advance and predict turning trajectories. An optimized computer vision detector enhanced with a high-frequency feature extraction branch enables robust small-target recognition for tower-crane viewpoints, ensuring stable long-term trajectory extraction. The proposed multimodal Transformer model incorporates dynamic localized environmental features that reflect road shape and obstacles, which strongly influence turning behavior on construction sites. Experiments using real construction video demonstrate that the method reduces mean trajectory prediction error by 12.2% relative to a trajectory-only Transformer baseline, with more than 14% improvement in turning scenarios. Furthermore, this study presents an end-to-end integration framework that maps predicted risk regions to physical warning devices in real time, enabling targeted, viewpoint-robust alarms without additional calibration even when camera viewpoints change. This study demonstrates that localized dynamic environmental information is a necessary predictor of nonlinear movement in field conditions and provides a deployable method for proactive collision prevention in dynamic construction sites, thus supporting the intelligent management of collision accidents at construction sites.

  • Open Access

    Training data generation for construction instance segmentation from target domains

    by Xuzhong Yan, Zeli Wang

    This study addresses a persistent challenge in computer vision for construction monitoring: deep learning models trained on source-domain data often perform poorly when deployed in new target domains due to distribution shifts and limited annotations. To mitigate these issues, the research introduces TDG-CIS, a clustering-initialized semi-supervised framework designed to generate high-quality instance segmentation training data directly from unlabeled target-domain images. TDG-CIS operates in two stages. First, it employs a clustering-based mask generation strategy that uses a transformer feature backbone to extract patch-level representations and derive initial instance masks without human supervision. These masks serve as a reliable starting point for semi-supervised learning. Second, a semi-supervised instance segmentation model iteratively refines these masks and converts raw images into usable training samples. This iterative pipeline allows the model to progressively improve segmentation quality while adapting to the visual characteristics of diverse construction environments. The framework was validated on a large dataset of 50,000 images spanning more than 70 construction-related domains. Experimental results show that TDG-CIS achieves a 77.9% data utilization rate, along with 87.5% mAP and 81.1% mAR in segmentation quality. When used to scale training data for downstream instance segmentation models, TDG-CIS yields substantial performance gains: baseline models trained on automatically generated data outperform those trained on manually labeled datasets, improving mAP from 92.9% to 94.3% and mAR from 86.7% to 88.6%. Ablation studies further demonstrate that the semi-supervised refinement mechanism is key to boosting both data utilization and segmentation accuracy. Overall, the study offers a novel approach that eliminates dependence on source-domain supervision and provides a scalable pathway for producing target-domain training datasets for instance segmentation in intelligent construction applications.

  • Open Access

    Intelligent Flood Scene Understanding Using Computer Vision-Based Multi-Object Tracking

    by Xuzhong Yan, Yiqiao Zhu, Zeli Wang, Bin Xu, Liu He, Rong Xia

    Understanding flood scenes is essential for effective disaster response. Previous research has primarily focused on computer vision-based approaches for analyzing flood scenes, capitalizing on their ability to rapidly and accurately cover affected regions. However, most existing methods emphasize static image analysis, with limited attention given to dynamic video analysis. Compared to image-based approaches, video analysis in flood scenarios offers significant advantages, including real-time monitoring, flow estimation, object tracking, change detection, and behavior recognition. To address this gap, this study proposes a computer vision-based multi-object tracking (MOT) framework for intelligent flood scene understanding. The proposed method integrates an optical-flow-based module for short-term undetected mask estimation and a deep re-identification (ReID) module to handle long-term occlusions. Experimental results demonstrate that the proposed method achieves state-of-the-art performance across key metrics, with a HOTA of 69.57%, DetA of 67.32%, AssA of 73.21%, and IDF1 of 89.82%. Field tests further confirm its improved accuracy, robustness, and generalization. This study not only addresses key practical challenges but also offers methodological insights, supporting the application of intelligent technologies in disaster response and humanitarian aid.


  • Open Access

    Privacy-preserving method for face recognition based on homomorphic encryption

    by Zhigang Song, Gong Wang, Wenqin Yang, Yunliang Li, Yinsheng Yu, Zeli Wang, Xianghan Zheng, Yang Yang

    In recent years, facial recognition technology has been widely adopted in modern society. However, the plaintext storage, computation, and transmission of facial data have posed significant risks of information leakage. To address this issue, this paper proposes a facial recognition framework based on approximate homomorphic encryption (HE_FaceNet), aimed at effectively mitigating privacy leaks during the facial recognition process. The framework first utilizes a pre-trained model to extract facial feature templates, which are then encrypted. The encrypted templates are matched using Euclidean distance, with the final recognition being performed after decryption. However, the time-consuming nature of homomorphic encryption calculations limits the practical applicability of the HE_FaceNet framework. To overcome this limitation, this paper introduces an optimization scheme based on clustering algorithms to accelerate the facial recognition process within the HE_FaceNet framework. By grouping similar faces into clusters through clustering analysis, the efficiency of searching encrypted feature values is significantly improved. Performance analysis indicates that the HE_FaceNet framework successfully protects facial data privacy while maintaining high recognition accuracy, and the optimization scheme demonstrates high accuracy and significant computational efficiency across facial datasets of varying sizes.


  • Open Access

    Computer Vision System for Multi-Robot Construction Waste Management: Integrating Cloud and Edge Computing

    by Zeli Wang, Xincong Yang, Xianghan Zheng, Daoyin Huang, Binfei Jiang

    Sorting is an important construction waste management tool to increase recycling rates and reduce pollution. Previous studies have used robots to improve the efficiency of construction waste recycling. However, in large construction sites, it is difficult for a single robot to accomplish the task quickly, and multiple robots working together are a better option. Most construction waste recycling robotic systems are developed based on a client-server framework, which means that all robots need to be continuously connected to their respective cloud servers. Such systems are low in robustness in complex environments and waste a lot of computational resources. Therefore, in this paper, we propose a pixel-level automatic construction waste recognition platform with high robustness and low computational resource requirements by combining multiple computer vision technologies with edge computing and cloud computing platforms. Experiments show that the computing platform proposed in this study can achieve a recognition speed of 23.3 fps and a recognition accuracy of 90.81% at the edge computing platform without the help of network and cloud servers. This is 23 times faster than the algorithm used in previous research. Meanwhile, the computing platform proposed in this study achieves 93.2% instance segmentation accuracy on the cloud server side. Notably, this system allows multiple robots to operate simultaneously at the same construction site using only a single server without compromising efficiency, which significantly reduces costs and promotes the adoption of automated construction waste recycling robots.


  • Open Access

    Vision-Based On-Site Construction Waste Localization Using Unmanned Aerial Vehicle

    by Zeli Wang, Xincong Yang, Xianghan Zheng, Heng Li

    In the context of construction and demolition waste exacerbating environmental pollution, the lack of recycling technology has hindered the green development of the industry. Previous studies have explored robot-based automated recycling methods, but their efficiency is limited by movement speed and detection range, so there is an urgent need to integrate drones into the recycling field to improve construction waste management efficiency. Preliminary investigations have shown that previous construction waste recognition techniques are ineffective when applied to UAVs and also lack a method to accurately convert waste locations in images to actual coordinates. Therefore, this study proposes a new method for autonomously labeling the location of construction waste using UAVs. Using images captured by UAVs, we compiled an image dataset and proposed a high-precision, long-range construction waste recognition algorithm. In addition, we proposed a method to convert the pixel positions of targets to actual positions. Finally, the study verified the effectiveness of the proposed method through experiments. Experimental results demonstrated that the approach proposed in this study enhanced the discernibility of computer vision algorithms towards small targets and high-frequency details within images. In a construction waste localization task using drones, involving high-resolution image recognition, the accuracy and recall were significantly improved by about 2% at speeds of up to 28 fps. The results of this study can guarantee the efficient application of drones to construction sites.


  • Single image super resolution via neighbor reconstruction

    by Zhihong Zhang, Chen Xu, Zhonghao Zhang, Guo Chen, Yide Cai, Zeli Wang, Heng Li, Edwin R. Hancock


  • Open Access

    Depth-based subgraph convolutional auto-encoder for network representation learning

    by Zhihong Zhang, Dongdong Chen, Zeli Wang, Heng Li, Lu Bai, Edwin R. Hancock