This thesis considers perception systems for urban environments. It focuses on the task of tracking dynamic objects and in particular on methods that can maintain the identities of targets through periods of ambiguity. Examples of such ambiguous situations occur when targets interact with each other, or when they are occluded by other objects or the environment.ududWith the development of self driving cars, the push for autonomous delivery of packages, and an increasing use of technology for security, surveillance and public-safety applications, robust perception in crowded urban spaces is more important than ever before. A critical part of perception systems is the ability to understand the motion of objects in a scene. Tracking strategies that merge closely-spaced targets together into groups have been shown to offer improved robustness, but in doing so sacrifice the concept of target identity. Additionally, the primary sensor used for the tracking task may not provide the information required to reason about the identity of individual objects. ududThere are three primary contributions in this work. The first is the development of 3D lidar tracking methods with improved ability to track closely-spaced targets and that can determine when target identities have become ambiguous. Secondly, this thesis defines appearance models suitable for the task of determining the identities of previously-observed targets, which may include the use of data from additional sensing modalities. The final contribution of this work is the combination of lidar tracking and appearance modelling, to enable the clarification of target identities in the presence of ambiguities caused by scene complexity. ududThe algorithms presented in this work are validated on both carefully controlled and unconstrained datasets. The experiments show that in complex dynamic scenes with interacting targets, the proposed methods achieve significant improvements in tracking performance.
展开▼
机译：本文考虑了用于城市环境的感知系统。它着重于跟踪动态对象的任务，尤其是可以通过模糊时间段保持目标身份的方法。当目标彼此交互或被其他物体或环境遮挡时，就会出现此类歧义情况。 ud ud随着自动驾驶汽车的发展，推动了自动交付包裹的推动以及技术的日益普及对于安全，监视和公共安全应用，在拥挤的城市空间中获得可靠的感知比以往任何时候都更为重要。感知系统的关键部分是理解场景中对象运动的能力。跟踪策略将间隔很近的目标合并成组已显示出改进的鲁棒性，但是这样做却牺牲了目标标识的概念。此外，用于跟踪任务的主要传感器可能不会提供推理有关单个对象的身份所需的信息。 ud ud这项工作有三大主要贡献。首先是3D激光雷达跟踪方法的发展，该方法具有增强的跟踪紧密间隔的目标的能力，并且可以确定目标身份何时变得模棱两可。其次，本文定义了适合于确定先前观察到的目标身份的任务的外观模型，其中可能包括使用来自其他传感方式的数据。这项工作的最终贡献是激光雷达跟踪和外观建模的结合，可以在场景复杂性造成歧义的情况下澄清目标身份。 ud ud这项工作中介绍的算法在经过仔细控制和不受约束的数据集上均得到了验证。实验表明，在具有交互目标的复杂动态场景中，所提出的方法在跟踪性能上取得了显着改善。
展开▼
