TY - CHAP
T1 - Selective frame analysis for efficient object tracking
T2 - balancing speed with accuracy in MOT systems
AU - Jung Shah, Yubraj
AU - Guo, Yi
AU - Park, Laurence A. F.
AU - Obst, Oliver
PY - 2025
Y1 - 2025
N2 - Applications such as autonomous driving and video surveillance rely on Multiple Object Tracking (MOT) technology to accurately identify objects in video data. Real-time MOT systems are often challenged for continuously improving computational efficiency while maintaining the acceptable level of accuracy. For every frame, advanced algorithms for detection and tracking is used to identify and track objects. However, it is still unclear if they should be used across all frames, or whether it would be better to use the algorithm only for selected frames. This is an empirical question, best answered by experimental research. Here, we explore how frame skipping during object detection impacts tracking accuracy and speed in real-time MOT systems. We examined the trade-off between skipping and tracking robustness for given MOT tasks. The consequences of frame skipping were evaluated using publicly available MOT datasets (KITTI, MOT16, MOT17 and MOT20) in different skipping image frequencies. Frame skipping allowed us to achieve a negligible drop in the MOTA and HOTA score while giving us a big 80% boost in speed over regular baseline configuration.
AB - Applications such as autonomous driving and video surveillance rely on Multiple Object Tracking (MOT) technology to accurately identify objects in video data. Real-time MOT systems are often challenged for continuously improving computational efficiency while maintaining the acceptable level of accuracy. For every frame, advanced algorithms for detection and tracking is used to identify and track objects. However, it is still unclear if they should be used across all frames, or whether it would be better to use the algorithm only for selected frames. This is an empirical question, best answered by experimental research. Here, we explore how frame skipping during object detection impacts tracking accuracy and speed in real-time MOT systems. We examined the trade-off between skipping and tracking robustness for given MOT tasks. The consequences of frame skipping were evaluated using publicly available MOT datasets (KITTI, MOT16, MOT17 and MOT20) in different skipping image frequencies. Frame skipping allowed us to achieve a negligible drop in the MOTA and HOTA score while giving us a big 80% boost in speed over regular baseline configuration.
KW - Computational Efficiency
KW - Frame Skipping
KW - Multiple Object Tracking (MOT)
KW - Selective Frame Skipping
KW - Tracking-by-Detection
UR - http://www.scopus.com/inward/record.url?scp=86000438810&partnerID=8YFLogxK
UR - https://go.openathens.net/redirector/westernsydney.edu.au?url=https://doi.org/10.1007/978-981-96-0692-4_20
U2 - 10.1007/978-981-96-0692-4_20
DO - 10.1007/978-981-96-0692-4_20
M3 - Chapter
AN - SCOPUS:86000438810
SN - 9789819606917
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 245
EP - 256
BT - Multi-disciplinary Trends in Artificial Intelligence: 17th International Conference, MIWAI 2024, Pattaya, Thailand, November 11-15, 2024, Proceedings, Part I
A2 - Sombattheera, Chattrakul
A2 - Weng, Paul
A2 - Pang, Jun
PB - Springer
CY - Singapore
ER -