To address the problem of limited single-camera views and insufficient feature fusion of target tracking, an online multi-target multi-camera tracking method with global face-pedestrian feature fusion is proposed. Firstly, by constructing a global face-pedestrian feature pool and adopting a global long-term identity management mechanism, real-time multi-camera information fusion and identity association are realized during single-camera tracking. Secondly, multi-category feature extraction and adaptive threshold adjustment are integrated, and a spatio-temporal-appearance joint constraint model is introduced to enhance the accuracy and robustness of multi-camera correlation. Finally, in order to verify the method, a multi-camera pedestrian tracking dataset covering different complexities is constructed. Experimental results show that compared with those of the comparison methods, the average number of ID switches (IDS) is reduced by more than 20 times, and the maximum number of ID F1 score (IDF1) is increased by about 30%. In the multi-camera tracking task, IDS in the multi-pedestrian scene is reduced from 132 to 126. Ablation experiments confirm the effectiveness of each module of the proposed method, and the real-time test shows that frames per second (FPS) of this method can reach 25.15.
在改进的FairMOT[11](on the Fairness of Detection and Re-Identification in Multiple Object Tracking)基础上,扩展检测头与特征通道,将单类别跟踪模型扩展为可独立跟踪人脸与行人的多类别模型,再通过联合匹配实现二者动态绑定。框架包含以下两个阶段。
1)类内独立跟踪。人脸跟踪分支:通过DLA-34[12]主干网络提取人脸特征,输出人脸框(为人脸数)及人脸特征,利用联合检测与嵌入(Joint Detection and Embedding, JDE)跟踪器生成人脸轨迹,为人脸ID。行人跟踪分支:使用共享的DLA-34主干网络提取行人特征,输出行人框(为行人数)及行人重识别特征,利用JDE跟踪器生成行人轨迹,为行人ID。
SPECKERA. OCMCTrack: online multi-target multi-camera tracking with corrective matching cascade[C]//Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway, USA: IEEE, 2024:7236-7244.
[4]
SPECKERA, BEYERERJ. ReidTrack: Reid-only multi-target multi-camera tracking[C]//Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway, USA: IEEE, 2023:5442-5452.
[5]
NGUYEND M H, HENSCHELR, ROSENHAHNB, et al. LMGP: lifted multicut meets geometry projections for multi-camera multi-object tracking[C]//Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2022:8856-8865.
[6]
HOUY Z, ZHENGL. Multiview detection with shadow transformer (and view-coherent data augmentation)[C]//Proceedings of the 29th ACM International Conference on Multimedia. New York, USA: ACM, 2021:1673-1682.
[7]
QUACHK G, NGUYENP, LEH, et al. DyGLIP: a dynamic graph model with link prediction for accurate multi-camera multiple object tracking[C]//Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2021:13779-13788.
[8]
HEY H, WEIX, HONGX P, et al. Multi-target multi-camera tracking by tracklet-to-target assignment[J]. IEEE Transactions on Image Processing, 2020,29:5191-5205.
[9]
YANGF, ODASHIMAS, YAMAOS, et al. A unified multi-view multi-person tracking framework[J]. Computational Visual Media, 2024,10(1):137-160.
[10]
LEEW Y, JOVANOVL, PHILIPSW. Multi-view target transformation for pedestrian detection[C]//Proceedings of the 2023 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops. Piscataway, USA: IEEE, 2023:90-99.
[11]
ZHANGY F, WANGC Y, WANGX G, et al. FairMOT: on the fairness of detection and re-identification in multiple object tracking[J]. International Journal of Computer Vision, 2021,129(11):3069-3087.
[12]
YUF, WANGD Q, SHELHAMERE, et al. Deep layer aggregation[C]//Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2018:2403-2412.