I-Corps: Control for Visual Scene Perception

拟议的I-Corps项目的更广泛的影响/商业潜力包括显著提高计算机视觉和移动机器人的实际应用的安全性和控制。在智能环境的最佳监控和控制中的第一个应用将影响许多行业的安全和生产力，例如视频监控，访问控制和智能建筑/城市。计算机视觉软件应用程序将通过预测附近行人、动物或其他车辆的行为来影响自动驾驶汽车等行业。自动视频处理和识别软件将通过视频分析软件提供的意图预测和异常检测，潜在地减少恐怖袭击或大规模枪击等灾难的影响。此外，在不同的应用中，所提出的解决方案可用于优化建筑物的能源效率并降低建筑物的运营成本。这个I-Corps项目进一步开发了一个自主视觉场景感知和反馈控制平台。自动实时感知和预测控制将通过引导人类操作员注意情况或自动应用传感器视野或可控环境条件（如照明或温度）的变化，大大提高几种潜在应用的安全性和效率。独特的深度学习贝叶斯优化框架不需要预先了解其部署的场景，而是随着时间的推移学习场景表示。该项目将机器学习、评估和控制系统联系在一起。概念验证测试已经在配备摄像头的实时车辆上成功完成，并在环路中使用深度学习来自主控制传感器的视野。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of the proposed I-Corps project consists of significantly increased safety and control for real-world applications of computer vision and mobile robotics. The first application in the optimal monitoring and control of smart environments will impact the safety and productivity of many industries, such as video surveillance, access control, and smart buildings/cities. Computer-vision software applications will impact industries such as autonomous automobiles by predicting actions of nearby pedestrians, animals, or other vehicles. Automated video processing and recognition software will potentially reduce the impact of catastrophes such as terrorist attacks or mass shootings via intent prediction and anomaly detection available through the proposed video analytics software. Additionally, in a different application the proposed solution may be used to optimize a building's energy efficiency and reduce building operating costs.This I-Corps project further develops a platform for autonomous visual scene perception and feedback control. Autonomous real-time perception and predictive control will dramatically increase the safety and efficiency across several potential applications by directing a human operator's attention to a situation or automatically applying changes to the sensor field of view or controllable environmental conditions such as lighting or temperature. The unique deep learning Bayesian optimization framework does not require prior knowledge of the scene in which it is deployed and instead learns the scene representation over time. The project ties together machine learning, estimation, and control systems. Proof of concept testing has been successfully completed in real-time vehicles equipped with cameras and using deep learning in the loop for autonomous control of the sensor field of view.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.