一句话：提出 TrackTest 框架，用约束优化生成行为多样的轨迹并做优先级排序，系统暴露多目标跟踪系统的失效。
多目标跟踪（MOT）是自动驾驶感知栈的核心组件，为下游决策提供状态估计；即便采用多传感器融合，MOT 系统仍会出现跟踪失效，可能引发严重安全事故。现有感知测试主要依赖人工标注的真实数据集，行为不均衡、扩展成本高，难以覆盖多样的动态交互。为此，作者提出自动化测试框架 TrackTest：将真实种子场景转化为具有挑战性的测试用例，把六种代表性驾驶模式形式化为基于约束的优化问题，从而生成物理可行且行为可控、多样的轨迹；同时引入轨迹优先级排序策略，识别最可能触发跟踪失效的高风险场景以降低计算开销。作者在六个先进的 MOT 系统上评估 TrackTest，结果表明该框架能有效生成行为多样的轨迹并揭示多种类型的跟踪失效。
Multiple Object Tracking (MOT) is a fundamental component of the perception stack in Autonomous Driving Systems (ADS), providing essential state estimation for downstream decision-making. Despite the adoption of multi-sensor fusion to enhance reliability, MOT systems remain susceptible to tracking failures, which can lead to catastrophic safety incidents. Current perception testing and evaluation predominantly rely on manually annotated real-world datasets, which are behaviorally imbalanced, prohibitively expensive to scale, and insufficient to cover diverse dynamic interactions. To bridge this gap, we propose TrackTest, an automated testing framework designed to systematically expose tracking failures in MOT systems. TrackTest transforms real-world seed scenarios into challenging test cases by synthesizing physically feasible and behaviorally diverse trajectories. We formulate six representative driving patterns as constraint-based optimization problems for behavior-controllable trajectory generation. To reduce computational overhead, we introduce a trajectory prioritization strategy that identifies high-risk scenarios most likely to trigger tracking failures. We evaluate TrackTest on six state-of-the-art MOT systems. Experimental results show that TrackTest effectively generate behavior-diverse trajectories and uncover diverse tracking failures.
@inproceedings{paperbot12389,
title = {TrackTest: Trajectory Generation and Prioritization for Testing Multiple Object Tracking System},
author = {Xinyu Gao and Haoxin Chen and Shuoxiao Zhang and Enyi Tang and Hengrui Xu and Haoxiang Tian and Minghui Wei and An Guo},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3837540}
}