The main question to be researched in this project is how to understand and distinguish normal vs. abnormal patterns of life in a city — the pulse of the city. The focus will be on the City of Pittsburgh, Allegheny County, and on developing the analytical tools to determine normal and abnormal traffic patterns into and out of the City.
Cities are increasingly equipped with low-resolution cameras. They are cheap to buy, install, and maintain, and thus are usually the choice of departments of transportation and their contractors. Pittsburgh or New York City have networks of hundreds of cameras. Video from some of these cameras is publicly accessible in real time. In this project, we addressed the problem of building a traffic model for parts of the roads visible from publicly accessible cameras. In particular, our end goal is to build a model capable of detecting different types of vehicles in images in various weather conditions and times of the day except night. Models learn different appearance of vehicles as seen from different viewpoints. A major difficulty with any type of analysis like this is the need for large amounts of training data. In our case, it is easy to collect unlabeled data from publicly available low-resolution low-framerate cameras in Pittsburgh or NYC (figure 1). Fig. 1: typical images from NYC cameras Some contractors from industry recently made substantial investments into the manual labelling of millions of cars. Such a large-scale approach allowed them to come up with a complex cascade detector built on hand crafted Haar-like image representation. They report reaching 98% precision - 98% recall point. In this work, we aim to achieve similar performance, but without the prohibitively expensive human labelling.
|email@example.com||Moura, Jose||CMU||PI||Faculty - Tenured|
|Final Report||pulse.pdf||July 17, 2018, 4:48 a.m.|
No match sources!