ProbFlow-Net: A probabilistic flow-aware and temporally consistent framework for reliable multi-object tracking in complex crowded scenes

Multi-object tracking (MOT) in crowded and dynamic environments presents persistent challenges due to occlusions, irregular motion, and detection uncertainty. This paper introduces ProbFlow-Net, a flow-guided and uncertainty-aware tracking-by-detection framework that formulates object association as a probabilistic inference problem jointly conditioned on optical flow, detection uncertainty, and temporal feature reliability. The Probabilistic Flow-Guided Association Network (PFAN) consists of five interrelated modules: Optical Flow Feature Extraction (OFFE) for motion-aware representations, Uncertainty-Guided Feature Modulation (UGFM) for adaptive reliability weighting, Temporal Flow Consistency Module (TFCM) to maintain temporal consistency, Probabilistic Association Network (PAN) estimating posterior-style association probabilities for reliable matching, and Dynamic Tracklet Reweighting (DTR) for identity stability. Using FlowNet-S and a fine-tuned YOLOv11 detector, ProbFlow-Net achieves consistent results across MOT17, MOT20, DanceTrack, and KITTI, with HOTA scores of 67.2, 66.9, 66.6, and 59.1, respectively. The results show improvements over BoT-SORT, FocusTrack, and BoostTrack++ in several benchmark settings, demonstrating the framework's robustness to heavy occlusion and motion ambiguity while maintaining online tracking efficiency. ProbFlow-Net provides a scalable framework for motion- and uncertainty-aware tracking, enhancing realistic multi-object tracking applications in autonomous driving, surveillance, and intelligent robotics.