This article presents an algorithm for video surveillance devoted to road traffic monitoring. The proposed approach starts from the raw video of the road and ends with an estimation of the velocity for each car appearing in the images. Here, we present a detailed analysis for every step of the proposed solution and an evaluation for the whole pipeline. Moreover, we will face different problems such as camera jittering or dynamic background.
Report and Slides
Demo Videos and Images
An example of how our tracker system works. It has some limitations when it comes to occlusions and shadows.
Another video example with the tracking method of Wang (which is inspired by deep learning).
We can see that Kalman filter does a pretty good job estimating the velocity of the vehicles, which remains most of the time stable. However, it is highly dependant on the segmentation that sometimes can be noisy or affected by shadows or occlusions. On the other hand, the SDA+NN is very good at tracking the objects, but as the bounding boxes constantly change the size and aspect ratio, the velocity estimation is more unstable (in one frame we have 55 kmh and in the next one 220 kmh, for example). Also, the problem of Wang method is that it is up to us to give the original bounding box of the object to track, which depends on the segmentation that can be noisy.
Authors and Contributors
This project has been developed by Adrià Ciurana (@adriaciurana), Guim Perarnau (@Guim3) and Pau Riba (@priba) as a project for the subject M4. Video Analysis Project of the Master in Computer Vision.