Answers are grounded in this paper and cite their sources.
Real-time crowd anomaly detection is an important capability for smart-city video analytics, publicspace management and emergency response. Surveillance cameras generate continuous streams in whichunusual events may be rare, short-lived, spatially localised and highly dependent on scene context. Examplesinclude sudden crowd dispersal, abnormal motion, falls, fights, unauthorised movement, dangerous congestionand objects or vehicles appearing in restricted pedestrian areas. Manual monitoring does not scale to largecamera networks, while conventional vision systems often struggle with occlusion, illumination changes,camera motion, scene variation and the open-ended nature of abnormal behaviour. This paper reviews deeplearning approaches to crowd and video anomaly detection and proposes a hybrid CNN–Vision Transformer–Bidirectional Long Short-Term Memory (CNN–ViT–BiLSTM) framework with adaptive attention and edge–cloud deployment. The review organises methods into handcrafted and reconstruction-based approaches,convolutional models, recurrent and temporal models, Transformer-based methods, graph and interactionmodelling, weakly supervised event detection, and edge-oriented architectures. It examines commonly useddatasets including UCSD Ped1/Ped2, CUHK Avenue, ShanghaiTech, UCF-Crime, XD-Violence and StreetScene, while emphasising that dataset labels, scene diversity and evaluation protocols differ substantially. Theproposed framework uses a CNN stem for local motion and appearance features, a ViT encoder for globalspatial context, adaptive spatial-temporal attention for selecting informative regions and time steps, and aBiLSTM for sequence modelling. A confidence-aware decision layer generates anomaly scores and eventcategories where supervision permits. Edge devices perform frame sampling, feature extraction and immediatescreening; cloud services support model management, long-term analytics, audit and controlled retraining. Areproducible evaluation protocol is specified using frame-level and video-level metrics, false-alarm rate,latency, throughput, energy, bandwidth and cross-scene generalisation. The framework is a research proposaland no empirical performance is asserted.
No takes yet. Share an insight, caveat, or question.
Roopa Shavi, Sanjeev Kulkarni (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: