Sequential adaptation framework for efficient detection and adaptive tracking of vulnerable road users in real-world scenarios

Vulnerable Road Users (VRUs), including pedestrians and cyclists, require reliable detection and tracking for intelligent transportation systems. This study presents a YOLOv8-Dynamic-SORT framework that combines accurate VRU detection, gradient-based sequential adaptation, and adaptive multi-object tracking. A comparative evaluation of Faster R-CNN, YOLOv5, YOLOv7, and YOLOv8 variants identifies YOLOv8x as the detector offering the most favorable accuracy-latency trade-off, achieving a precision of 0.763, a recall of 0.485, an [email protected] of 0.541, and an F1-score of 0.593 at an inference latency of 9.9 ms per image. Faster R-CNN attains a higher recall (0.670) and F1-score (0.604) but requires 220 ms per image, which precludes real-time onboard deployment. To examine adaptation across domains, the VisDrone-trained detector is sequentially adapted to the Caltech Pedestrian dataset and the effect of the optimizer on knowledge retention is quantified on the original task. SGD-based adaptation reduces the forgetting measure on VisDrone2019 from $$0.097 \pm 0.007$$ (Adam) to $$0.021 \pm 0.003$$ and improves backward transfer from $$-\,0.097 \pm 0.007$$ to $$-\,0.021 \pm 0.003$$ , retaining a VisDrone [email protected] of $$0.519 \pm 0.008$$ after adaptation, while also reaching a higher Caltech [email protected] of $$0.608 \pm 0.03$$ compared with $$0.461 \pm 0.03$$ for Adam adaptation. For tracking, the proposed Dynamic-SORT module selects among five BoxMOT trackers at runtime according to measurable scene descriptors. Compared with the best fixed configuration (BoTSORT: HOTA 47.8, MOTA 68.9, IDF1 78.9), Dynamic-SORT attains HOTA 48.6, MOTA 69.4, and IDF1 80.1 at an additional decision overhead of 1.8 ms, while achieving an end-to-end throughput of 48 FPS. These results indicate that optimizer choice measurably affects knowledge retention during sequential adaptation, and that scene-conditioned tracker selection improves identity preservation at modest computational cost.

Authors

Institutions

Publication Details

Journal
Discover Artificial Intelligence
Published
2026-11-17
DOI
https://doi.org/10.1007/s44163-026-02414-6
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Sequential adaptation framework for efficient detection and adaptive tracking of vulnerable road users in real-world scenarios

Salman Liaquat, Faryal Aurooj Nasir
3 links
Discover Artificial Intelligence
Advanced Neural Network Applications
article

Sequential adaptation framework for efficient detection and adaptive tracking of vulnerable road users in real-world scenarios

Salman Liaquat, Faryal Aurooj Nasir
article en

Abstract

Vulnerable Road Users (VRUs), including pedestrians and cyclists, require reliable detection and tracking for intelligent transportation systems. This study presents a YOLOv8-Dynamic-SORT framework that combines accurate VRU detection, gradient-based sequential adaptation, and adaptive multi-object tracking. A comparative evaluation of Faster R-CNN, YOLOv5, YOLOv7, and YOLOv8 variants identifies YOLOv8x as the detector offering the most favorable accuracy-latency trade-off, achieving a precision of 0.763, a recall of 0.485, an [email protected] of 0.541, and an F1-score of 0.593 at an inference latency of 9.9 ms per image. Faster R-CNN attains a higher recall (0.670) and F1-score (0.604) but requires 220 ms per image, which precludes real-time onboard deployment. To examine adaptation across domains, the VisDrone-trained detector is sequentially adapted to the Caltech Pedestrian dataset and the effect of the optimizer on knowledge retention is quantified on the original task. SGD-based adaptation reduces the forgetting measure on VisDrone2019 from $$0.097 \pm 0.007$$ (Adam) to $$0.021 \pm 0.003$$ and improves backward transfer from $$-\,0.097 \pm 0.007$$ to $$-\,0.021 \pm 0.003$$ , retaining a VisDrone [email protected] of $$0.519 \pm 0.008$$ after adaptation, while also reaching a higher Caltech [email protected] of $$0.608 \pm 0.03$$ compared with $$0.461 \pm 0.03$$ for Adam adaptation. For tracking, the proposed Dynamic-SORT module selects among five BoxMOT trackers at runtime according to measurable scene descriptors. Compared with the best fixed configuration (BoTSORT: HOTA 47.8, MOTA 68.9, IDF1 78.9), Dynamic-SORT attains HOTA 48.6, MOTA 69.4, and IDF1 80.1 at an additional decision overhead of 1.8 ms, while achieving an end-to-end throughput of 48 FPS. These results indicate that optimizer choice measurably affects knowledge retention during sequential adaptation, and that scene-conditioned tracker selection improves identity preservation at modest computational cost.

Discover Artificial IntelligenceVol. 6(1)
Universiti Sains Malaysia (MY), Institute of Space Technology (PK), National University of Sciences and Technology (PK)
Sustainable cities and communities, Industry, innovation and infrastructure
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon