Traffic safety margin inference via machine learning for 3D spatial modeling

点击翻译按钮获取中文摘要

A great deal of road and traffic intersection safety is based on statistics of accidents, injuries, and fatalities, but which leads to the dreadful conclusion that a statistically significant number of accidents, injuries, and fatalities need to take place before changes and improvements can begin to be made.Surely it would be enormously preferable to observe intersections and to make note of close calls (near-accidents) which are far more common than actual accidents, but which do permit assessments of safety to be estimated. Although the human eye and visual system are fairly good at spotting near-accidents, it is unrealistic to place people, for days or weeks, at uncountable intersections around the country.Instead, many cities and municipalities have already installed video systems, built by Miovision Technologies Inc. of Kitchener, Ontario, which could provide the relevant data. Currently such data are used for counting vehicles of different types, as a function of date and time, for traffic planning purposes. However the large existing installation of these video systems offers an outstanding opportunity to provide a significant value-added in terms of safety assessment.The University of Waterloo is proposing to partner with Miovision to develop advanced methods of computer vision to measure the nearest distance as a vehicle passes a pedestrian or a bicyclist, to obtain data on close calls which far exceed the 364 cyclist and pedestrian deaths per year, creating a valuable product for Miovision, essential data for municipalities and planners, and ultimately improved quality of life for Canadians.