About

The Human Detection in Video (HaDiVe) project is a culmination of work undertaken by faculty and students at the NYU Center for Urban Science + Progress. Using publicly available video streams the HaDiVe project aims to create a real-time pedestrian counter within New York City. At present, counts are recorded at approximately one minute intervals for each camera. Ultimately, this is being accomplished using a Faster Region-based Convolutional Neural Network (Faster R-CNN) trained with a hand-labelled set of DOT images. Visit the active GitHub repo.

Click on the map for more info.

Examples

With the current training and test set, the Faster R-CNN is achieving 51% precision identifying people. Together, the images below illustrate the current successes and shortcomings of the current iteration. As seen, the model is confidently identifying individuals in the foreground of the images. However, individuals who are distant from the camera or partially occluded are not being identified. Expanding the current training set is likely to improve overall performance. Nonetheless, current results may inform the relative number of pedestrians throughout the city at an granular temporal scale.


Download

All data from the previous month is available for download.

Prof. Gregory Dobler, Trang Dam, and Jordan Vani

Center for Urban Science + Progress

2017 NYU CUSP. All rights reserved.