Automatic Detection of Suspicious Behavior at Railway Crossings

Abstract

In this thesis we study the feasibility of a system to automatically detect suicidal and generally suspicious behavior from surveillance camera feeds in order to assist in the prevention of suicides at railway crossings. We design a system that is able to detect several types of behavior by using computer vision techniques such as object detection and multiple object tracking. We present an approach to extract behaviors from detected person trajectories and to classify these as being suspicious or not, with the intent of sending alerts to camera operators to notify them about railway crossings that need their attention. The aim is to assist these camera operators in their surveillance work. We implement and evaluate the performance of our system and discuss potential areas where improvements to our system can be made.