Counting and tracking people from image sequences

Current methods in automatic traffic monitoring are not suitable for counting pedestrians and cyclists. For example, inductive loops are based on detecting ferromagnetic masses and machine vision applications usually assume that objects are rigid and their trajectories are known in advance.

In case of the light traffic, i.e. pedestrians and cyclists, problems arise because of the small amount of structural constraints, occlusions and non-predictable trajectories. Also portability of the counting device causes some extra requirements. The solution that we have developed for this problem is based on machine vision. Using a simple motion detection algorithm for a digitized image sequence we extract regions of activity from the static background.

We also have to infer the number of the persons in each of these regions. In this stage we apply snake-technique for finding the boundaries of the objects in the region. The number of persons can be now obtained by following the shape of the snake-curve.

This procedure is repeated for each frames in the sequence and the positions of the detected objects and the corresponding time stamps are stored. By analysing the data afterwards we can count the total number of persons passed by during the time camera has been measuring. We also get trajectories and velocities of these people so that we can distinguish between pedestrians and cyclists.

An experimental system is implemented in workstation environment using Matlab. Despite of the low frame rate (2 frames/s) the counting accuracy is good. With the groups less than four persons the accuracy is better than 90%.

An example

By pressing this button you can see the original image sequence captured from the laboratory window. There are two people walking side by side in the street. Now, let's look at what the computer sees: Yellow rectangles are the regions of activity and red curves are the snakes fitted to those regions. Green circles show the position of detected persons. After processing the sequence we obtain a figure of the trajectories. By examining this data computer can easily decide that there has been two persons in the sequence.

Click this figure to zoom the trajectories.