Detection of Complex Events under Uncertainty in a Video Surveillance System

Detecting complex crowd events in video surveillance is challenging due to visual uncertainty, occlusions, and ambiguous group dynamics.This paper proposes a hybrid neuro-symbolic framework that integrates Mask R-CNN for instance-level detection and feature extraction with a fuzzy ontology for uncertainty-aware semantic reasoning.The fuzzy ontology models imprecise concepts via membership functions and SWRL rules, enabling robust inference of high-level events-walking, running, formation, splitting, evacuation, and dispersion-from noisy lowlevel observations.Evaluated on the PETS 2012 dataset, the framework achieves an average precision of 83 %, recall of 87 %, and F-measure of 85 %.An ablation study quantifies the contribution of the fuzzy ontology, showing a 12 % average F-measure gain over a Mask R-CNN baseline with crisp post-processing.Additional experiments on the UCF-Crime dataset (81 % F-measure for 'Fighting', 79 % for 'Arrest') confirm strong generalization.The work demonstrates that fuzzy ontology serves as an effective, interpretable interface between statistical deep learning and symbolic event reasoning, offering a principled solution for complex event detection under uncertainty.