ZZZhixiang Zhu

Zhixiang Zhu

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zhixiang Zhu has a total of 13 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Research on video synopsis optimization method based on Two-stage abnormal behavior recognition

    by Fudan Wang, Zhixiang Zhu, Chenwu Wang, Pei Wang, Lei Zhang, Pengyuan Mu

    Video synopsis techniques shift the moving target in the time axis to achieve the purpose of displaying as much effective information as possible in the shortest possible time, usually the abnormal behavior in the video contains more effective information, and most of the existing video synopsis methods do not consider the processing of abnormal behavior. To solve this problem, this paper proposes a video synopsis method based on two-stage abnormal behavior recognition. First, the abnormal frames containing abnormal behaviors in surveillance video are detected using a spatiotemporal autoencoder method, and then an improved R(2+1)D network model is used to identify the abnormal targets and abnormal behavior classes in the abnormal frames of surveillance video. Finally, the abnormal priority concentration loss term is added to the video synopsis optimization objective loss function to make the abnormal targets appear more preferentially and centrally in the generated synopsis video by the multi-objective optimization algorithm. Through a large number of experimental results, the method proposed in this article not only prioritizes the appearance of abnormal targets in the synopsis video, but also accurately observes the specific abnormal behavior of abnormal targets.

  • Open Access

    SR-VIReID: Super Resolution Assisted Visible-Infrared Person Re-Identification

    by Shaobo Hei, Zhixiang Zhu, Chenwu Wang, Pei Wang, Linfeng Yang

    Visible-infrared person re-identification (Re-ID) aims to match the pedestrian images of the same identity from different modalities. Existing works mainly focus on alleviating the modality discrepancy by aligning the distributions of features from different modalities. However, these methods usually assume that the gallery image and the query image have the same resolution and sufficient fine-grained detail, the accuracy of these methods for Re-ID at low resolutions leaves much to be desired. To address this problem, we propose a super-resolution-assisted infrared visible pedestrian Re-ID method (SR-VIReID) in this paper. Firstly, based on the MPANet, the backbone is improved by using CSPNet to extract low-level texture features and high-level semantic features by means of stacked CSP, CBR, and SPP structures Secondly, a simple SR branch is introduced to learn high-resolution (HR) feature representation, which SRs pedestrians in low-resolution images to obtain more detailed information and further improve re-recognition accuracy. In order to avoid extra computational effort, the SR module is removed in the inference stage. Experimental results show that, on SYSU-MM01 and RegDB datasets, the method in this paper improves the Re-ID accuracy to some extent without adding extra computation.


  • Open Access

    Interaction preservation based Multi-feature fusion for video synopsis generation

    by Kaixuan Yang, Zhixiang Zhu, Chenwu Wang, Pei Wang

    Faced with a large number of surveillance videos, it will cause users to spend a lot of time browsing and retrieving videos, Video synopsis technology solves this problem, which condenses the spatiotemporally redundant long video into a compact summary video and preserves the activity information of all objects. However, the spatial and temporal displacement of objects during the synopsis process destroys the interaction between objects. This paper proposes a video synopsis method that preserves the interaction between objects. First, the depth prediction of the objects in the video is performed, and then the surveillance video is subjected to inverse perspective Mapping (IPM) to calculate the distance between objects, Finally, the depth and distance are used as the factors for judging the interaction relationship of objects, and the interaction relationship between the objects in the original video and the synopsis video is kept consistent by optimizing the interaction energy cost term. Experimental results show that the proposed method effectively preserves the interaction between objects in the synopsis.


  • Open Access

    Joint Multi-Scale Residual and Motion Feature Learning for Action Recognition

    by Linfeng Yang, Zhixiang Zhu, Chenwu Wang, Pei Wang, Shaobo Hei

    For action recognition, two-stream networks consisting of RGB and optical flow has been widely used, showing high recognition accuracy. However, optical flow computation is time-consuming and requires a large amount of storage space, and the recognition efficiency is very low. To alleviate this problem, we propose an Adaptive Multi-Scale Residual (AMSR) module and a Long Short Term Motion Squeeze (LSMS) module, which are inserted into the 2D convolutional neural network to improve the accuracy of action recognition and achieve a balance of accuracy and speed. The AMSR module adaptively fuses multi-scale feature maps to fully utilize the semantic information provided by deep feature maps and the detailed information provided by shallow feature maps. The LSMS module is a learnable lightweight motion feature extractor for learning long-term motion features of adjacent and non-adjacent frames, thus replacing the traditional optical flow and improving the accuracy of action recognition. Experimental results on UCF-101 and HMDB-51 datasets demonstrate that the method proposed in this paper achieves competitive performance compared to state-of-the-art methods with only a small increase in parameters and computational cost.