City Scape Camera develops deep learning methodologies to extract vehicle counts from streaming real-time video.
Our methodologies are specially adapted to low resolution web cameras and we focused on publicly available cameras installed in the Manhattan borough of New York City.

Our goal is contribute to the study of urban mobility, particularly the analysis of traffic patterns as seen by a network of cameras. We do this using density-based counting methods that are highly precise, yet privacy friendly.

We created a large annotated dataset that we make available to the whole community. See description below and details for license and download instruction in the DATASET page

Scientific Contributions

- A multitask CNN based architecture to count cars by mapping images to a density-image.
- An LSTM based architecture for high precision counting exploring space-time correlations
- An adversarial domain adaptation strategy to handle images from cameras where there isn't training data available.

This work was supported and partially funded by the Portuguese Science Foundation FCT under the framework of the Carnegie Mellon|Portugal Program. For full details see our publications page.

CityCam Dataset

In the scope of our research, CSCam created a unique dataset that we make publicly available. The dataset has the following features:

- 60,000 annotated frames
- 900,000 annotations of objects: vehicle count, type, bounding box, re-id, orientation, timestamp, weather
- 60,000,000 test images

This dataset, owned by CMU, is made publicly available under the licence and user agreement described in the DATASET page.

US Patent Application

This work contributed to the PCT Patent Application No:PCT/US18/26341

Title: Deep Learning Methods for Estimating Density and/or Flow of Objects, and Related Methods and Software
Filed on May 4th, 2018
Inventors : José F. Moura, João Paulo Costeira, Shanghang Zhang and Evgeny Toropov
Applicants: Carnegie Mellon University and Instituto Superior Técnico