Izvorni znanstveni članak
https://doi.org/10.7307/ptt.v37i2.635
An M-DeepSORT Algorithm for Pedestrian Detection and Tracking Based on Video Images – A Case Study in Ji-nan Subway Station
Wei ZHANG
; Beijing Jiaotong University, School of Systems Science
Chuang ZHU
; Beijing Jiaotong University, School of Systems Science
*
Yunchao QU
; Beijing Jiaotong University, School of Systems Science
Guanhua LIU
; Beijing Jiaotong University, School of Systems Science
Der-Horng LEE
; Zhejiang University, Zhejiang University; University of Illinois at Urbana-Champaign (ZJU-UIUC) Institute
* Dopisni autor.
Sažetak
With the ongoing urbanisation, the subway has become a vital component of modern cities, catering to the escalating demands of a mobile population. However, the increasing complexity of passenger flows within subway stations poses challenges to operations management. To optimise subway operations and enhance safety, researchers focus on extracting and analysing pedestrian trajectories within subway stations. Traditional trajectory extraction methods face limitations due to manual feature design and multi-stage processing. Leveraging advancements in deep learning, this paper integrates M-DeepSORT with YOLOv5 and proposes a feature association matching approach that addresses trajectory drift issues through simultaneous consideration of motion and appearance matching. The confidence-based (CB) Kalman filtering method is proposed to address the issue of random noise in pedestrian detection within subway scenes. The introduction of a momentum-based passenger trajectory centre update method reduces jitter, resulting in smoother trajectory extraction. Experimental results affirm the effectiveness of the proposed algorithm in detecting, tracking and statistically analysing subway station corridor passenger flow trajectories, demonstrating robust performance in diverse subway station scenarios.
Ključne riječi
passenger trajectory tracking; CB Kalman filtering; trajectory update; momentum; M-DeepSORT
Hrčak ID:
329163
URI
Datum izdavanja:
30.4.2025.
Posjeta: 0 *