Boosting Open-Vocabulary Multiple Object Tracking With Wavelet Convolution and Confidence-Aware Kalman

Open-vocabulary multi-object tracking (MOT) aims to generalize trackers to object categories unseen during training. Existing methods mainly rely on appearance features extracted by conventional convolutions, which neglect high-frequency motion details and low-frequency structural information in images, making them insufficient for robust perception. Meanwhile, in open-vocabulary scenarios, detection confidence is degraded and measurement noise is amplified, leading to the failure of traditional Kalman filter–based modeling. To address these limitations, we propose OVSORTv2. Specifically, OVSORTv2 tackles the instability in tracking low-confidence targets and the limited representation ability of conventional convolutions by three key improvements: the Confidence-Weighted Kalman update (CWK) module, which dynamically adjusts filter weights according to detection confidence; the Confidence-Noise Adaptive Kalman (CNAK) module, which enhances motion modeling through confidence-noise coupling; and the Haar Wavelet Convolution (HWTConv) module, which employs wavelet decomposition and frequency-wise convolutions to enlarge the receptive field and strengthen the representation of motion and appearance features. Extensive experiments on the TAO validation and test sets demonstrate the superiority of OVSORTv2, showing improved robustness in dynamic open scenarios and stronger adaptability to novel categories, compared to existing methods.