Department of Electrical and Electronic Engineering University of Cagliari

Watch the demo

Trajectory classification

The trajectory of a moving object (a person, a car, etc.) can be labelled as being "typical" or "atypical" with respect to the characteristics of the suirveilled scene.

As an example, in a parking area, typical trajectories of cars are those that follow the rules, otherwise they are considered as being "atypical".

Our approach involves four steps:

STEP 1 - Data collection

Collection of a number of "typical" trajectories. These trajectories may be either recorded in a live scenario, or they can be performed by some actors, or they can be synthetised according to the rules of the considered application. In the figure, a number of synthetic trajectories of cars and pedestrians in a car parking scenario. These trajectories are considered as being "typical" for the considered scenario.

STEP2 - Trajectory representation

Each trajectory has been represented by means of a Bezier curve. In the figure below, the original trajectory is displayed in red, while the corresponding representation by a Bezier curve is displayed in blue.

STEP 3 - Feature extraction

Each trajectory is sampled in a number of points at equal distance. For each point, the (x,y) coordinates are stored as well as the local slope dx/dy. These data make up the reference data for classifying unknown trajectories.

STEP 4 - Classification

An unknown trajectory is first represented by a Bezier curve, and a number of points are sampled according to the rules followed at Step 3. This trajectory is then classified as being atypical if its distance by the nearest typical trajectory is higher than the distance between the two nearest typical trajectories. Otherwise the trajectory is classified as beying typical. The distance between two trajectories is computed as the Euclidean distance in the space made up of 3*p components, where p is the number of points sampled for each trajectory, and 3 refers to the three information related to each point (x,y,dx/dy). It is worth noting that data related to atypical trajectories is not needed.

Input data for the demonstrator

The input data for the demonstrator to be developed is made up of

- a set of trajectories represented in terms of a sequence of (x,y) coordinates. They represent a set of typical trajectories for the considered scenario

- the trajectory that should be labelled according to the representative data.