Minimax-Based Intelligent Multimodal Surveillance for Real-Time Threat Detection: A Structured Review

Abstract

Surveillance systems now generate far more video than any control room can watch, and the research response has been to add modalities. Audio, pose, optical flow, depth, language, hyperspectral bands and trajectory data have all been folded into threat-detection pipelines under the single banner of multimodal surveillance. This review argues that the banner conceals four largely separate research programmes with different assumptions, different benchmarks and different notions of what counts as success. It reads twenty-one studies published in IEEE journals between 2025 and 2026 against that split, grouping them by modality composition, by where fusion happens, and by supervision regime. Three findings come out of that reading. Audio, the modality most often invoked to justify the label, is genuine in only one public benchmark, and several leading systems synthesise it, which makes the reported audio contribution partly an artefact of the synthesiser. Real time is claimed by most of the reviewed work but measured at incompatible boundaries, with feature extraction excluded from the latency budget often enough that cross-paper timing comparison is close to meaningless. And modality degradation, the failure mode that determines whether a fused system survives deployment, is tested by two studies in twenty-one. The review organises the corpus into three comparative tables and closes on research gaps written as small concrete experiments, not as a call for a unified architecture.