Abstract
The presented project aims to solve the problem of road traffic monitoring. The main goal is concerned about counting the amount of vehicles on a road and estimate their speed using visual cues. Our solution shows an optimized approach through the typical adaptive Gaussian modelling in order to subtract the background. Then, some well-known morphological filters and shadow removal techniques are applied to refine the proposed methodology. Moreover, stabilization methods have been studied to eliminate jitter effect caused by wind. Finally, a Kalman filter is used to track each vehicle appearing in the sequence with a homography matrix that obtains the correct perspective for computing the final speed.
Interested researchers can find more information and details about the aforementioned approach by clicking this link.
Supplementary material
This section includes the slides for each week deliverable, explaining in an accurate way the methodology that has been followed.
Week 1: Introduction.
Week 2: Background extraction.
Week 4: Camera stabilization.
Week 5: Tracking and speed estimation.
In addition, we have made public and available through GitHub the whole Matlab code regarding our implementation. Please, we will appreciate if you cite our work in your publication as:
M. Alfaro, A. Calafell, M. Matilla and J. Torrents, "Beyond a Smart Video Surveillance approach for Road Traffic Monitoring", 3rd Workshop on Road Traffic Monitoring, vol. 1, pp. 1-5, 2016.
Demonstration
The next recorded video sequences present exactly how our tracker system works. Notice that, our final approach uses an adaptive method for background estimation, showing its efficacy in challenging sequences as Traffic. Although, we expected to obtain better results using Gaussian Mixture Model, in especially when background pixels can take different values.
TRAFFIC Sequence:
HIGHWAY Sequence:
OUR Sequence:
It is important to take into account that our system is able to deal with some common issues, like camera stabilization. However, others must be improved as shadows removal (particularly if they are as dark as cars color) and the speed computation when the camera is well-oriented.
In addition, we have compared our results with some state-of-the-art publications. Concretely, below we have added a resulting video sequence inspired on a Deep learning approach made by Wang et al. (2013). The presented tracker uses unsupervised feature learning in order to perform the vehicle tracking. Then, a stacked denoising autoencoder is pre-trained offline using natural images obtaining robust generic features against variations.
Authors & Developers
This project has been developed by the next multidisciplinary team: Mónica Alfaro, Andrea Calafell (LinkedIn), Martín Matilla (LinkedIn) and Jordina Torrents, as a project for the module M4 - Video Analysis of the Master in Computer Vision.