Detecting human motion with support vector machines

This work presents a method for detection of humans in video sequences. The intended application of the method is outdoor surveillance. In such an uncontrolled environment, the appearance of humans varies hugely due to clothing, identity, weather and amount and direction of light. The idea is therefore to detect patterns of human motion, which to a large extent is independent of the differences in appearance. To this end, a support vector machine is trained with dense optical flow patterns originating from humans. The subjects are moving in different angles to the camera plane, on different image scales. This trained SVM is the core of a human detection algorithm which searches optical flow images for human-like motion patterns.

1. Introduction

Detection of humans in image sequences is an active research area within computer vision. In many applications, such as human-computer interaction, the detection is a basis for tracking. In other, such as surveillance systems or safety systems in cars, the detection in itself is used to trigger some type of alarm. The intended application of the method presented in this paper is outdoor surveillance.