I-Corps: Real-Time Traffic Congestion Detection from Surveillance Videos

这个i-Corps团队建议开发一种从监控视频中实时检测交通拥堵的系统。拟议的交通拥堵检测技术涉及两个主要部分：(A)车辆检测和跟踪；以及(B)事件分类。研究小组将使用计算机视觉技术在单帧图像中检测车辆。通过分析连续的帧，可以估计出车辆的速度和相对位置。对于事件的分类，团队将利用他们在图形核和机器学习方面的专业知识。可以使用一系列连续的帧来创建图形，其中节点表示用诸如速度和位置等局部特征标记的车辆。图中的相邻节点由标有各自车辆之间距离的边连接。图的快速核函数将由二进制分类器开发和使用，该分类器将被训练用于识别交通拥堵的任务。作为扩展，还可以训练多类分类器来区分不同类型的交通事件，如高、中、低交通拥堵、事故或正常交通。所提出的产品是将最先进的计算技术转化为可以直接产生社会和商业影响的新技术的一个很好的机会。实时交通拥堵检测系统以全国各地的政府机构和部门为主要目标。有了实时拥堵检测器，人们将能够同时跟踪数百或数千个摄像头，发现可以加强交通管理的事件。决策者可以将建议的产品用于动态交通分配、事件发现和改进的疏散系统管理。通勤时间和延误成本的节省可能会对整个社会产生影响。该产品还可以用于开发网络应用程序或移动设备应用程序，为驾车者带来实时交通信息。

This I-Corps team proposes the development of a real-time traffic congestion detection system from surveillance videos. The proposed technology for detecting traffic congestion involves two major components: (a) vehicle detection and tracking; and (b) event classification. The research team will use computer vision techniques for detecting vehicles in single frames. By analyzing consecutive frames, vehicle speed and relative locations will be estimated. For the classification of events, the team will use their expertise in graph kernels and machine learning. A sequence of consecutive frames can be used to create a graph, where the nodes represent vehicles labeled with local features, such as speed and location. Neighboring nodes in the graph are connected by edges labeled with the distance between their respective vehicles. A fast kernel function for graphs will be developed and used by a binary classifier which will be trained for the task of recognizing traffic congestion. As an extension, a multi-class classifier can also be trained to distinguish different types of traffic events, such as, high, moderate, or low traffic congestion, accident, or normal traffic.The proposed product is a great opportunity for transforming state-of-the-art computational techniques into new technologies that can directly have societal and commercial impacts. A real time traffic congestion detection system has as their primary targets government agencies and departments across the nation. With a real time detector of congestion, one will be able to track simultaneously hundreds or thousands of cameras at the same time, discovering incidents that can enhance traffic management. Decision makers can use the proposed product for dynamic traffic assignment, incident discovery, and improved management of evacuation systems. Society may be impacted as a whole with savings in commuting time and delay costs. This product can also be used for developing web applications or apps for mobile devices, bringing real time traffic information to motorists.