Background subtraction with deep learning

点击翻译按钮获取中文摘要

Osprey Informatics Inc. has a large clientele from Oil and Gas sector interested in automated surveillance of outdoor scenes. Manual monitoring is tedious, error prone and expensive. Thus, Osprey Informatics is looking out to build automated and computer vision-based solutions. The very first step in a computer vision-based automated monitoring of a video stream is known as background subtraction (BGS), which finds out motion and temporal changes that occur in the scene being imaged. Once BGS is performed, further processing and analysis of the scene can be accomplished toward automated surveillance. Existing solutions for BGS have been proved ineffective for the challenging data sets Ospery deal with. Challenges within the data sets include long time lapse between consecutive images, outdoor scene changes due to natural lighting variations, appearance of shadows, and camera shakes due to windy conditions. The goal of this research is to build a BGS solution, which is able to mitigate the effects of the aforementioned adverse factors as much as possible by using the latest neural network-based tools from computer vision.