Using Shape, Motion and Context for Object Classification in 3D Point Clouds

自动驾驶的关键要素之一是对运动物体的鲁棒和精确检测、跟踪和分类。在之前的研究中，主要是根据物体的形状和运动来进行分类。对于这个任务，包含形状和运动信息的对象模型是基于手工标记的轨迹创建的。在这个后续项目中，这些模型应该被自动学习。为了将物体划分为不同的物体类别，同时提高对运动的预测，需要考虑物体的形状、运动和上下文。人类以一种非常自然的方式使用上下文。它可以被描述为一组从经验中产生的隐含规则。例如，行人在人行道上行走，人行道在马路旁边，汽车在马路上行驶。在本研究中，应该开发一种不使用固定规则而自动学习对象上下文的方法。预期的好处是即使在困难的情况下也能对移动物体进行稳健的分类和预测。这种困难情况的例子是，如果存在自遮挡或来自其他物体的遮挡，则形状信息不完整；如果观察到的物体不移动，则缺少运动信息。背景与形状和运动的额外组合有望提高对运动物体的鲁棒分类和预测。

One of the key elements of autonomous driving is the robust and precise detection, tracking and classification of moving objects. In the previous research on this topic, the shape and the movement of an object was used for classification. For this task, object models containing shape and movement information were created based on hand labeled tracks. In this follow-up project, these models should be learned automatically. To separate the objects into different object classes and to improve the prediction of the movement at the same time, the shape, the movement and the context of an object should be taken into account. Humans use the context in a quite natural way. It can be described as a set of implicit rules generated from experience. For example, pedestrians move on sidewalks, sidewalks are beside the roadway and cars move on the roadway. In this research, a method should be developed that learns the context of an object automatically without the use of fixed rules. The expected benefit is the robust classification and prediction of moving objects even in difficult situations. Examples of such difficult situations are incomplete shape information if self-occlusion or occlusion from other objects is present and missing movement information if an observed object is not moving. The additional combination of the context with the shape and the movements is expected to improve robust classification and the prediction of moving objects.