Zero-Shot TrackingT0：对象分割+运动感知记——当“切万物”武士学会运动记忆，目标跟踪稳如老狗

💻 实战效果演示

🔧 核心机制

1. 运动建模系统（Motion Modeling）

▍原SAMv2缺陷

▍SAMURAI新方案

class KalmanEnhancedSelector:def __init__(self):self.kf = KalmanFilter(dim_x=8, dim_z=4)  # 状态向量[x,y,w,h,dx,dy,dw,dh]self.alpha = 0.2  # 论文验证def select_mask(self, masks: List[Mask], scores: List[float]) -> Mask:""" 融合运动与外观的双路径决策 """kf_predictions = [self.kf.predict() for _ in masks]kf_scores = [iou(kf_pred, mask.bbox) for mask, kf_pred in zip(masks, kf_predictions)]combined_scores = [self.alpha * kf_score + (1-self.alpha) * mask_score for kf_score, mask_score in zip(kf_scores, scores)]return masks[np