Training-Free Vision–Language Framework for Sidewalk Pedestrian Behavior Recognition
Publication Type
Conference Paper
Authors
Fulltext
Download

Understanding how pedestrians use sidewalks is important for urban mobility, safety, and effective pedestrian-
space management. Automatic behavior recognition can help characterize these usage patterns. However, conventional approaches often require task-specific annotated data and training, while labeled sidewalk-behavior datasets remain limited. We propose a training-free Vision–Language framework combining Grounding DINO for pedestrian region extraction and Ristretto-3B for behavior classification. Evaluation was conducted on a new, manually annotated test set of 215 instances across eight behavior classes. For behavior classification, Ristretto-3B achieves 91.10% macro accuracy and 91.36% F1-score, outperforming two baselines. The complete framework achieves 84.31% macro accuracy and 82.59% F1-score. Grounding DINO improves end-to-end macro accuracy by 11.29 percentage points over YOLO-11x. Qualitative examples illustrate prompt-based vocabulary expansion and task reconfiguration without task-specific retraining.

 
Conference
Conference Title
International Conference on FutureTech 2026 (ICFT)
Conference Country
Jordan
Conference Date
Oct. 18, 2026 - Oct. 22, 2026
Conference Sponsor
IEEE
Additional Info
Conference Website