Feature Extraction

Introduction

Primarily we have used movement based features, which seek to capture the characteristics of the trajectory of a moving object in the video. Thus movement based features should be able to discriminate between a fast moving roller skater (anomaly) and slow moving pedestrians (non-anomaly). We can also augment these features with shape based features, which may be useful to identify anomalies like carts and vehicles, which clearly have a different shape than pedestrians.

Features

Assuming non-anomalous behaviour is people simply walking, movement characteristics are discriminative between anomalous and non anomalous objects because:

• Many anomalies are due to fast moving objects like cars, cycles or people on roller skates. Hence speed is clearly a good feature.
• People tend to walk in a roughly constant speed. Hence acceleration is an indicator that something anomalous happened, for example a person starts running which may be suspicious behaviour.
• People also tend to walk in a fairly straight line, while some anomalous objects like a roller skating person tends to follow a meandering trajectory. Hence features that characterize the shape of the path is useful.

Based on these criteria, we have the following feature vector that captures the above mentioned criteria.

Let X denote the vector of x coordinates of a N point trajectory, and Y be the y coordinates. Let Xd be the pairwise differences between the elements of X, that is a discrete approximation of the derivative of X. Similarly we define Yd for the Y vector. Let Vp = sqrt((Xd)² + (Yd)²) be the pairwise velocity magnitudes. Let Ap be the pairwise acceleration vector that we get by performing pairwise difference on Vp. Let D be the vector containing the deviations from the straight line connecting end points of the trajectory for each point