Towards a new benchmark for background subtraction algorithms in computer vision

Abstract

Background subtraction is the initial step in many computer vision applications. As such, choosing the most appropriate algorithm for a given task is crucial as it can significantly impact the performance of the entire system. However, the selection process is often challenging due to the large amount of available algorithms, and the lack of standardised benchmarks. Traditionally, researchers and industry professionals have relied on paper surveys and tools like ChangeDetection.net (CDNet) to identify potential algorithms. This thesis attempts at addressing these limitations by proposing a novel Background Subtraction Recommender Platform, a scalable, extendable, and modular web-based system that adapts to the user's input to recommend the most suitable algorithms for his needs, abstracting away the complexity of the selection process. The main goal of the application is to provide a contextualised ranking of the algorithms, and to streamline the algorithm development process by allowing the submission and execution of new algorithms while providing insights on their performances. To do so, this solution leverages state-of-the-art algorithm evaluation procedures. Additionally, contextualisation of the recommendation process is guided by semantic segmentation algorithms to put emphasis on the content of the video sequences. The ranking system is evaluated for a set of algorithms obtained from the BGSLibrary on real-world data built upon the well-known CDNet dataset, demonstrating its ability to provide meaningful recommendations to the user.