A Fuzzy Logic-Based System for Soccer Video Scenes Classification

Massive global video surveillance worldwide captures data but lacks detailed activity information to flag events of interest, while the human burden of monitoring video footage is untenable. Artificial intelligence (AI) can be applied to raw video footage to identify and extract required information and summarize it in linguistic formats. Video summarization automation usually involves text-based data such as subtitles, segmenting text and semantics, with little attention to video summarization in the processing of video footage only. Classification problems in recorded videos are often very complex and uncertain due to the dynamic nature of the video sequence and light conditions, background, camera angle, occlusions, indistinguishable scene features, etc. Video scene classification forms the basis of linguistic video summarization, an open research problem with major commercial importance. Soccer video scenes present added challenges due to specific objects and events with similar features (e.g. “people” include audiences, coaches, and players), as well as being constituted from a series of quickly changing and dynamic frames with small inter-frame variations. There is an added difficulty associated with the need to have light weight video classification systems working in real time with massive data sizes. In this thesis, we introduce a novel system based on Interval Type-2 Fuzzy Logic Classification Systems (IT2FLCS) whose parameters are optimized by the Big Bang–Big Crunch (BB-BC) algorithm, which allows for the automatic scenes classification using optimized rules in broadcasted soccer matches video. The type-2 fuzzy logic systems would be unequivocal to present a highly interpretable and transparent model which is very suitable for the handling the encountered uncertainties in video footages and converting the accumulated data to linguistic formats which can be easily stored and analysed. Meanwhile the traditional black box techniques, such as support vector machines (SVMs) and neural networks, do not provide models which could be easily analysed and understood by human users. The BB-BC optimization is a heuristic, population-based evolutionary approach which is characterized by the ease of implementation, fast convergence and low computational cost. We employed the BB-BC to optimize our system parameters of fuzzy logic membership functions and fuzzy rules. Using the BB-BC we are able to balance the system transparency (through generating a small rule set) together with increasing the accuracy of scene classification. Thus, the proposed fuzzy-based system allows achieving relatively high classification accuracy with a small number of rules thus increasing the system interpretability and allowing its real-time processing. The type-2 Fuzzy Logic Classification System (T2FLCS) obtained 87.57% prediction accuracy in the scene classification of our testing group data which is better than the type-1 fuzzy classification system and neural networks counterparts. The BB-BC optimization algorithms decrease the size of rule bases both in T1FLCS and T2FLCS; the T2FLCS finally got 85.716% with reduce rules, outperforming the T1FLCS and neural network counterparts, especially in the “out-of-range data” which validates the T2FLCSs capability to handle the high level of faced uncertainties. We also presented a novel approach based on the scenes classification system combined with the dynamic time warping algorithm to implement the video events detection for real world processing. The proposed system could run on recorded or live video clips and output a label to describe the event in order to provide the high level summarization of the videos to the user.