Project description

This page serves as an overview of the project developed over the course of M4. The goal of the project is to count moving cars in a highway and estimate their speed. To do so we have explored the well known approach of background modelling to retrieve the moving objects. Using morphological operations we have been able to substantially improve both the numerical and visual performances of our system. We conducted a study of stabilization to improve the image quality reducing the effect of wind and pulse that were adding a jitter to the recordings. Finally, we developed from zero a Kalman filter to estimate the position and speed over the road of each car. The result is an autonomous system that can be used for road surveillance and traffic monitoring in outdoors environments.

Material

All the code can be found at the github page, in addition we have written a paper, and some slides.

Custom implementations

During the development of the project, and as the only team of 2017 implementing our system in Python, we found several function implementations that were easily done in MATLAB but did not have a good Python counterpart. Some of these alternatives presented additional issues that hintered our progress; difficult installations, lack of a multiplatform implementation or complete documentation. We made the best effort to try to overcome these obstacles, and for the use of the next master students, we have created a small list with all the functionalities we could not find in OpenCV or standard Python libraries:

• Kalman Filter : we implemented our own.
• MATLAB function bwareaopen : we implented our own.
• Hole filling : we implemented our own with the help of scipy-image library.
• Lucas Kanade : although OpenCV offers a sparse LK version, it lacks a dense counterpart. Moreover, there are no more Dense Optical Flow implementations in OpenCV, apart from Farneback.

Despite these lacks, we enjoyed developing our code in Python, as we believe is a more robust and accesible language, aimed at a wider range of tasks than MATLAB. The access to some of its powerful libraries made some of the very needed non computer vision tasks such as managing files or communicating with the Operative System much easier. Needless to say, we are open to point at the specific code locations where these problems were solved.

Authors and contributors

• Axel Barroso (@axelBarroso)
• Sergio Sancho (@sergiosanchoasensio)
• Alejandro Nespereira (@alejandronespereira)
• Marc Carné (@marccarne)