FlowSORT: a semisupervised optical flow-enhanced framework for robust multi-object tracking

We introduce FlowSORT, a semi-supervised optical flow-enhanced framework for robust multi-object tracking in real-time traffic monitoring. Compared with existing approaches, FlowSORT integrates a novel iterative optical flow quality refinement module—leveraging expert voting and probabilistic uncertainty estimation—with a flow-guided Kalman filter and an improved YOLOv7-based detector enhanced by channel attention. Trained on both KITTI and Sintel datasets, FlowSORT not only generates high-quality optical flow annotations for unlabeled field data but also corrects labeling errors in non-target domain datasets. Our framework achieves state-of-the-art performance on the MOT 17 benchmark, obtaining an mAP of 89.71% and reducing the optical flow endpoint error from 2.86 to 1.24, while enhancing target localization accuracy by 24.3%. In addition, after applying TensorRT quantization, FlowSORT operates 2.8× faster than baseline methods, maintaining a stable processing latency of 28 ms. These results demonstrate that FlowSORT offers a highly efficient and accurate solution for intelligent traffic monitoring and smart city applications.