EVENT DOSSIER
The researchers proposed the TrajMind framework for quickly and accurately diagnosing anomalies in collective trajectories in cities. This framework employs a combination of slow and fast methods: the slow path (TrajMind_slow) sequences steps such as type recognition, conditional localization, and verification on a frozen visual-linguistic backbone network to generate structured evidence; the fast path (TrajMind_fast) performs window screening through a single pure-text scan. Experimental results show that TrajMind_slow outperforms the strongest baseline by 15.3 percentage points in type recognition, 13.8 percentage points in localization accuracy, and maintains stable performance during cross-city migrations; meanwhile, TrajMind_fast reduces latency by 41.1% and maintains an accuracy rate of over 93.5% for binary balance.
TrajMind 提出一种快慢结合框架，通过切换三个角色专用 LoRA 适配器在冻结的视觉 - 语言骨干网络上实现城市轨迹集体异常诊断。其慢路径 TrajMind$_{\text{slow}}$串联基于画布的类型识别、序列化轨迹的条件定位及可执行验证，生成结构化证据；快路径 TrajMind$_{\text{fast}}$以单次纯文本扫描进行窗口筛查。实验表明，TrajMind$_{\text{slow}}$在异常类型识别上比最强基线高 $15.3$ 个百分点，在定位上高 $13.8$ 个百分点，且在跨城市迁移中保持性能；TrajMind$_{\text{fast}}$将延迟降低 $41.1\%$ 并保持二元平衡准确率不低于 $93.5\%$。