Visual Models - Application to Situational Awareness

我的研究解决了计算机视觉表示的问题，特别是如何从传感器网络捕获信息，并创建随着时间的推移忠实地反映环境的计算机模型。本提案中有3个具体的子项目。第一个涉及为与人类一起工作的机器人助手提供情景感知。对象（人类+机器）被表示为铰接的3D模型，其在真实的时间中跟踪其真实的世界对应物。科学的重点是如何捕捉和建模动态行为，以便系统可以识别或控制每个占用者的活动。在之前与通用汽车研究合作的基础上，该公司的目标是将机器人助手嵌入更强大、更可靠的情景感知，以便实现实际应用。如果这一愿景成功，未来的装配线将由人类-机器人团队合作制造产品，大大提高生产力，并使加拿大在全球市场上保持竞争力。 剩下的两个子项目旨在使人类能够更精确地了解其环境，并使机器人能够在其中发挥作用。在第一个稀疏数据模型中，我们研究了如何在有限的感知能力下恢复环境描述的问题。具体的例子是从物理岩心样本中定位地下矿藏，这些样本相对于图像中发现的样本来说非常稀疏。我们最近已经开发了一些新的随机建模技术，这些技术有望将传统的图像重建算法扩展到这些具有挑战性的数据集。如果成功，这项研究将导致更精确的算法来定位矿床，这反过来可能通过降低挖掘，运输和加工成本对采矿业务产生重大影响。 最后，深度学习子项目试图利用机器学习中令人印象深刻的技术进步来确定更适合自然形式的表示。我们的重点是使用卷积深度玻尔兹曼机网络的分层生成模型的实际实现，目标是在大幅降低复杂性的情况下实现相当的性能。这项研究将有助于与通用汽车公司合作，建立可以在恶劣天气条件下在越野环境中运行的传感器系统，以及涉及在GPS缺失环境中识别结构和地标的资源行业项目。

My research addresses the problem of representations for computer vision, specifically how to capture information from a sensor network and create computer models that faithfully reflect an environment over time. There are 3 specific sub-projects within this proposal. The first deals with providing situational awareness for a robotic assistant working in conjunction with a human. Objects (humans + machines) are represented as articulated 3D models that track their real world counterparts in real time. The scientific focus is how to capture and model dynamic behavior so that the system can identify or control the activities of each occupant. Following on previous work with GM research, the goal is to embed robotic assistants with more powerful and reliable situational awareness so that practical implementations become possible. If this vision is successful, future assembly lines will be comprised of human-robot teams collaborating to manufacture products, greatly scaling up productivity and allowing Canada to remain competitive in the global marketplace. The remaining 2 sub-projects are aimed at representations that enable humans to have more precise knowledge about their environments, and robots to be able to function within them. In the first, Sparse Data Models, we look at the problem of how to recover descriptions of an environment with limited sensory ability. The specific example is localizing underground mineral deposits from physical core samples, which are very sparse relative to what one finds in images. We recently have developed some new stochastic modeling techniques that show promise for extending conventional image reconstruction algorithms to these challenging datasets. If successful, this research will lead to more precise algorithms for localizing mineral deposits, which in turn could have a significant impact on mining operations by reducing costs for excavation, transport and processing. Finally, the Deep Learning sub-project is an attempt to leverage impressive technical progress in machine learning to determine representations that are better suited to natural forms. Our focus is on the practical implementation of Hierarchical Generative Models using Convolutional Deep Boltzman machine networks, with the goal of achieving comparable performance with a substantial reduction in complexity. This research will contribute to work with GM in building sensor systems that can operate in off-road environments under adverse weather conditions, as well as resource industry projects that involve identifying structures and landmarks in GPS-deprived environments.