Video cropping using salience maps: a case study on a sidewalk dataset

Abstract

Video cropping aims trim video frames to highlight a subject area. This paper introduces a new framework for automated video cropping tailored to sidewalk footage, which is particularly useful in applications such sidewalk navigability and urban planning. By developing a method for video salience annotation using simple mouse input, the introduced framework provides a simple and flexible approach for video cropping. This application is crucial in scenarios where accurately focusing on pedestrian areas is necessary to enhance analysis and decisionmaking processes. The experimental results obtained from real data in the wild shows that the method is robust to a large variety of sidewalk conditions in different Brazilian cities.