Video analytics for abnormal event detection

点击翻译按钮获取中文摘要

The goal of this project is to design new methods to detect abnormal events in videos and to generate semantic labels to describe objects in videos. For abnormal event detection, we will focus on the detection of the loitering behavior, which consists in spending a large amount of time in a given area. Since the loiterer can exit and re-enter the field of view of the camera, we will also need to investigate person re-identification. The loitering detection is meant to be run on each camera to lighten the image processing load at the main visual surveillance hub. Therefore, the scientific challenges will be to differentiate individuals that may have similar appearance, and track the individuals, while at the same time designing methods that can be implemented on an embedded system. To solve this problem, we will rely on machine learning techniques to learn to compare the appearance of objects. This will be useful for both tracking and person re-identification. ****We will also study the problem of generating semantic labels to describe the individuals that are tracked. This is to accelerate search in videos. We wish to generate labels that describe the general appearance of the person, like the color of their clothes, their type (skirt, pant, etc.), if they are wearing a hat or carrying a bag. To solve this problem we will learn to generate labels from the appearance of the individuals in the image. The learned appearance to semantic labels neural network will also be useful for person re-identification.