Xudong Fan

Dr. Xudong Fan

Beihang University

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

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

Representative Publications

View All Publications
24 February 2022 in IEEE Transactions on Intelligent Transportation Systems
Verified
Journal Article
Closed access
Xudong FanXudong Fan
Xudong Fan
WZWei Zhao
Wei Zhao
Automatic license plate recognition plays an important role in intelligent transportation systems and is of great significance. However, at present, most current approaches are only concerned of license plate recognition under restrictive conditions, where the license plates are shot in a frontal view and under good light conditions. These approaches are not robust enough in real-world complex capture scenarios, such as uneven light condition or oblique shooting angle. In order to improve the robustness of recognizing license plates under complex capture scenarios, a robust license plate detection network (CA-CenterNet) is proposed in this paper, together with a segmentation-free network (CNNG) for the recognition of license plate characters. CA-CenterNet can detect not only the center of each license plate, but also four vectors pointing to the four corners of the corresponding license plate, regardless of the rotation and distortion of the license plates, which gives us the possibility to rectify the distorted license plates in the source images. Then, CNNG can accurately identify the characters in the detected license plates without character segmentation. Experimental results prove that our automatic license plate recognition system has good performance in real-world complex capture scenarios and outperforms current license plate recognition models.

15 November 2020 in Sensors
Verified
Journal Article
Open access
PDF file
Ruihong YinRuihong Yin
Ruihong Yin
WZWei Zhao
Wei Zhao
Xudong FanXudong Fan
Xudong Fan
Yongfeng YinYongfeng Yin
Yongfeng Yin
There are a large number of studies on geospatial object detection. However, many existing methods only focus on either accuracy or speed. Methods with both fast speed and high accuracy are of great importance in some scenes, like search and rescue, and military information acquisition. In remote sensing images, there are some targets that are small and have few textures and low contrast compared with the background, which impose challenges on object detection. In this paper, we propose an accurate and fast single shot detector (AF-SSD) for high spatial remote sensing imagery to solve these problems. Firstly, we design a lightweight backbone to reduce the number of trainable parameters of the network. In this lightweight backbone, we also use some wide and deep convolutional blocks to extract more semantic information and keep the high detection precision. Secondly, a novel encoding–decoding module is employed to detect small targets accurately. With up-sampling and summation operations, the encoding–decoding module can add strong high-level semantic information to low-level features. Thirdly, we design a cascade structure with spatial and channel attention modules for targets with low contrast (named low-contrast targets) and few textures (named few-texture targets). The spatial attention module can extract long-range features for few-texture targets. By weighting each channel of a feature map, the channel attention module can guide the network to concentrate on easily identifiable features for low-contrast and few-texture targets. The experimental results on the NWPU VHR-10 dataset show that our proposed AF-SSD achieves superior detection performance: parameters 5.7 M, mAP 88.7%, and 0.035 s per image on average on an NVIDIA GTX-1080Ti GPU.

11 September 2025 in IEEE Transactions on Geoscience and Remote Sensing
Verified
Journal Article
Closed access
HCHaoxiang Chen
Haoxiang Chen
WZWei Zhao
Wei Zhao
Xudong FanXudong Fan
Xudong Fan
XSXiping Shang
Xiping Shang
RZRufei Zhang
Rufei Zhang
DLDongjin Li
Dongjin Li
Data annotation is extremely difficult due to the satellite imaging conditions, under which the targets are usually small, obscured, and scattered. Therefore, satellite video data annotation inevitably has noise and errors. Moreover, the irregular acceleration, sudden turns, and stops of the target make prediction and trajectory maintenance highly challenging. In this study, we propose the Points of Interest Flows Track (PIFTrack) to address the aforementioned challenges. PIFTrack improves tracking accuracy by modeling target uncertainty distributions and nonlinear motion patterns, while leveraging the spatial inclusion relationships of points of interest (PoIs) across consecutive frames. Specifically, we eliminate the rigid Dirac-based labeling assumption by employing a set of PoIs to model the spatial probability distribution of the target. PoIs enable the model to infer optimal outputs in the vicinity of annotations, thereby improving robustness to annotation errors. Secondly, to capture the real motion transfer patterns of targets in the data, we introduce a diffusion-based ordinary differential equation (ODE) model. Ultimately, we alleviate the impact of tiny object localization drifts on association results by exploiting the inclusion relationship between PoIs. PIFTrack has been extensively evaluated on the VISO, AIR-MOT, CGSTL, and VSMB datasets, exhibiting competitive performance relative to contemporary studies. Our code is open-source and available at https://github.com/softwarePupil/PIFTrack.

18 December 2025 in IEEE Transactions on Geoscience and Remote Sensing
Verified
Journal Article
Closed access
XSXiping Shang
Xiping Shang
WZWei Zhao
Wei Zhao
HCHaoxiang Chen
Haoxiang Chen
Xudong FanXudong Fan
Xudong Fan
NLNannan Li
Nannan Li
RZRufei Zhang
Rufei Zhang
Few-shot object detection for remote sensing images (RSIs) has received much attention. The main difficulty in few-shot object detection currently lies in the imprecise classification, which can be tackled from three key perspectives: features to be classified (Region of Interest features), reference features (feature prototypes), and classifiers. However, existing methodologies have not comprehensively addressed these aspects. To fill this gap, we propose the Unconstrained Feature Enhancement Text-Guided Few-Shot Remote Sensing Image Object Detector (UFEDet). Initially, we introduce the text-image shared-specific module, which not only focuses on the consistent representation of the same class across different modalities but also emphasizes the unique information of each modality, thereby enhancing the distinctiveness of class feature prototypes. Subsequently, we propose an unconstrained feature enhancement module, which enhances the Region of Interest (RoI) features by leveraging the global class information from feature prototypes, while also increasing the distributional differences between classifiers. Finally, we introduce a gradient-controlled optimizer that modulates the gradient based on the gradient response of base classes, alleviating catastrophic forgetting of base classes during the training of novel classes. To validate the efficacy of our proposed approach, experiments were conducted on the DIOR and NWPU VHR-10 datasets. The experimental results demonstrate that the mAP of both base and novel classes exceeds that of popular few-shot object detectors.

1 December 2019
Verified
Conference Paper
Open access
Xudong FanXudong Fan
Xudong Fan
DBDaniele Bonatto
Daniele Bonatto
GLGauthier Lafruit
Gauthier Lafruit
Face swapping in videos usually has strong entertainment applications. Deep Fakes (in Faces) are a recent topic in deep learning where the main idea is to substitute the face of a person in a video with the face of another person. But one of the drawbacks of the method is that between two successive frames there are inconsistencies between the faces, such as changing face color, flickering or eyebrows that change. In this paper, we propose a convolutional neural network for swapping faces based on two autoencoders which share the same encoder. In this network, the encoder can distinguish and extract important features of faces, including facial expressions and poses; the decoders will then reconstruct faces according to these features. First, we will generate datasets of faces respectively for person A and person B. Secondly, the local information of two faces is sent to the network to get the model; after the training process, we can use the model to reconstruct the corresponding face of person B when the input is one face of person A. Afterwards, we build a binary mask to select the face area and transfer color from the source face to the target face. Finally, we only need to use a seamless clone to merge the new faces back into the source frames to create a fake video. The experimental results show that the quality of the fake videos is improved significantly.