SBIR Phase I: Large-Scale Behavioral Analysis Utilizing Convolutional Neural Networks and Its Application to In-store Retail Marketing

小型企业创新研究(SBIR)第一阶段项目的更广泛的影响/商业潜力是开发一个基础设施，用于暴露和解释以前无法获得的数据集：细粒度的人类与物理环境的交互。人类在不断地建设和塑造世界，但几乎没有数据来检验这些影响。除了零售业，这项技术还可能影响教师布置教室的方式、灾难工作人员提供救援的方式，或者工厂保护工人安全的方式。每天影响人类的微妙物理细节将被理解和调查，如果没有拟议的系统，这是不可能的。这项技术将特别通过提高零售商的经济效益以及广泛地通过增加对人类如何与其物理环境交互的科学理解来造福社会。这项小型企业创新研究(SBIR)第一阶段项目使用神经网络和通用3D场景重建，并结合低成本的无线摄像头来建模具有准确的对象分类和空间关系的环境。最近，神经网络已被证明擅长于各种图像分类任务，但它们在视频分类中的应用，即针对人类行为的应用，还很少被探索。3D场景重建也取得了类似的进展，从图像到视频，但总是需要一些物理场景的先验知识。在过去的一年里，多个小组提出了从多个视角的摄像机重建完全通用的场景的方法。最后，相机和无线发射器等设备的能量收集方法已在实验室实验中证明是可行的，但尚未纳入商业产品。在这个项目中，两种计算机视觉算法将与摄像机硬件并行开发，以便在项目结束时可以集成和演示软件和硬件系统。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to develop an infrastructure for exposing and interpreting a previously unavailable dataset: fine-grained human interaction with a physical environment. Humans are continuously building and shaping the world but there exists little data to examine these effects. Beyond retail, this technology could affect how teachers layout classrooms, how disaster workers provide relief, or how factories keep their workers safe. The subtle physical details that affect humans everyday will be understood and investigated in ways not possible without the proposed system. This technology will benefit society specifically by improving the economic efficiency of retailers and broadly by increasing scientific understanding of how humans interact with their physical environments.This Small Business Innovation Research (SBIR) Phase I project uses a neural network and generic 3D scene reconstruction in combination with low-cost, wireless cameras to model an environment with accurate object classification and spatial relationships. Recently, neural networks have proven adept at a variety of image classification tasks but their applications in video classification, namely for human actions, have been less explored. 3D scene reconstruction has made similar advances, progressing from images to videos but has always required some prior knowledge of the physical scene. Within the past year, multiple groups have proposed methods for completely generic scene reconstruction from multiple view cameras. Finally, energy harvesting methods for devices such as cameras and wireless transmitters have been demonstrated to be feasible in laboratory experiments but have not been incorporated into commercial products. In this project the two computer vision algorithms will be developed in parallel with camera hardware so that the software and hardware systems may be integrated and demonstrated by the end of the project.