While data augmentation serves as a potent strategy to encompass corner cases in autonomous driving, scant research has attended to the imperative of maintaining categorical balance post-augmentation. Hereby, We introduce a novel data augmentation technique designed to proficiently regulate the fine-grained distribution of pedestrian poses within datasets, thereby amplifying the efficacy of pedestrian recognition algorithms in traffic situations. Our approach is founded on a unique quadratic categorization of human pose situations, using corresponding pose estimation confidence scores. This methodology strives to precisely adjust the proportion of poses within differing confidence score ranges. Experimental outcomes demonstrate that our approach not only enhances the distribution of pedestrian poses in traffic image datasets, but also significantly boosts the performance of pedestrian detectors in final.


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Data Augmentation for Pedestrians in Corner Case: A Pose Distribution Control Approach


    Beteiligte:
    Nie, Yunhao (Autor:in) / Chen, Yuanyuan (Autor:in) / Miao, Qinghai (Autor:in) / Lv, Yisheng (Autor:in)


    Erscheinungsdatum :

    24.09.2023


    Format / Umfang :

    2006599 byte





    Medientyp :

    Aufsatz (Konferenz)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch




    Intention Recognition of Pedestrians and Cyclists by 2D Pose Estimation

    Fang, Zhijie / López, Antonio M. | ArXiv | 2019

    Freier Zugriff

    Data Augmentation for RFID-based 3D Human Pose Tracking

    Wang, Ziqi / Yang, Chao / Mao, Shiwen | IEEE | 2022


    Pedestrians

    TIBKAT | 2018


    Pedestrians

    TIBKAT | 2016