Advancing Object Detection and Tracking Frontiers in Intelligent Vision-Based Applications

目标检测和跟踪（ODT）被认为是大多数智能视觉应用的基石。我们将对象检测定义为一个广泛的研究领域，它可以识别某些类别的对象的存在（众所周知的分类任务）并定位它们的位置。对象跟踪识别视频或图像序列中相同对象随时间的轨迹。智能视觉应用市场预计将从2021年的122亿美元增长到2027年的205亿美元。许多领域都依赖于这些基于智能视觉的应用，包括无人机视觉、智能视频监控、自动驾驶、医疗和健康应用以及安全。最近，最先进的ODT模型在大量标记训练数据集的帮助下使用监督学习取得了巨大成功，在某些情况下甚至超过了人类水平的性能（例如ImageNet上的分类）。然而，这些模型在它们可以解决的问题的范围方面仍然有限，它们需要“增加它们的域外鲁棒性”。“换句话说，他们在特定领域的专业任务上表现良好，但当领域发生变化时，他们“在狭窄的领域或范围之外往往很脆弱，”正如图灵奖获得者Bengio，LeCun和欣顿最近在2021年7月的一篇文章中指出的那样。适应领域的变化对人类来说是很自然的，但对于基于智能视觉的应用程序来说仍然是一个巨大的挑战。最近的一个悲惨的现实例子是2018年3月的致命碰撞，这是由于自动驾驶汽车的视觉系统在夜间将行人错误分类为在不同帧中以不同速度移动的不同类别的物体（未知物体，然后是车辆，最后是自行车）。因此，在不同的场景和班次下，ODT性能的高度准确性和一致性至关重要。通过这样做，这将有助于我们朝着开发可靠的智能系统的方向发展，这些系统可以像人类一样学习和适应。该研究计划的长期目标是通过开发和创建下一代目标检测和跟踪模型来推进基于智能视觉的应用。更具体地说，在短期内，在未来五年内，我将解决以下两个主题：1）开发对象检测中的域泛化新技术，2）开发对象跟踪中的鲁棒跨域外观模型新技术。该研究计划将为至少13名HQP提供培训，帮助他们在图像处理，计算机视觉，深度学习，优化和计算复杂性分析等高级主题方面建立强大的背景。

Object detection and tracking (ODT) is considered a cornerstone in most intelligent vision-based applications. We define object detection as the broad research area that identifies the presence of objects of certain classes (the well-known Classification task) and localizes their position. Object tracking identifies the trajectories of the same objects over time in a video or a sequence of images. The intelligent vision-based applications market is expected to grow from $12.2 billion in 2021 to $20.5 billion in 2027. Many domains rely on these intelligent vision-based applications, including drone vision, intelligent video surveillance, autonomous driving, medical and health applications, and security. Recently, state-of-the-art ODT models have achieved great success using supervised learning with the aid of massive labelled training datasets, even surpassing human-level performance in some cases (e.g. classification on ImageNet). However, these models are still limited in terms of the scope of the problems they can solve, and they need to "increase their out-of-domain robustness." In other words, they perform well on the specialized tasks in the specific domains they have been trained on (in-distribution), but when a domain shift happens, they "are often brittle outside of the narrow domain or scope they have been trained on," as noted in a recent July 2021 article by the Turing Award winners Bengio, LeCun, and Hinton. Adapting to domain shifts is natural to humans but is still a massive challenge for intelligent vision-based applications. A recent tragic real-life example is a March 2018 fatal collision resulting from the vision system of a self-driving car miss-classifying a pedestrian for whole six seconds during the night as different classes of objects moving at different speeds in different frames (unknown object, then as a vehicle, and finally as a bicycle). Hence, it is critical for ODT performance to be both highly accurate and consistent under different scenarios and shifts. In doing so, this will help us to progress towards developing reliable, intelligent systems that can learn and adapt much like humans. The long-term objective of this research program is to advance intelligent vision-based applications through developing and creating the next-generation object detection and tracking models. More specifically, in the short term, over the next five years, I will address the following two themes: 1) developing new techniques for domain generalization in object detection and 2) developing new techniques for robust cross-domain appearance models in object tracking. This research program will provide training to at least 13 HQP, helping them build strong backgrounds in advanced topics in image processing, computer vision, deep learning, optimization, and computational complexity analysis.