ANOMALY DETECTION AND LOCALIZATION IN CROWDED SCENES

Crowded scene surveillance has become an essential component in ensuring public safety, smart city
monitoring, transportation security, and emergency management. Detecting abnormal activities in dense crowds
is a challenging computer vision problem due to severe occlusions, dynamic motion patterns, illumination
changes, and the unpredictable nature of anomalies. Traditional surveillance systems largely depend on human
operators, making the process labor-intensive, error-prone, and inefficient for real-time applications.
This paper presents an intelligent anomaly detection and localization framework for crowded scenes using deep
learning and probabilistic motion modelling techniques. The proposed system combines Mixture of Dynamic
Textures (MDT) for modelling normal crowd motion and Conditional Random Fields (CRF) for accurate
spatial–temporal anomaly localization. Deep convolutional feature extraction and temporal sequence learning
modules further improve the robustness of abnormal event recognition.
The framework processes live or recorded surveillance video, extracts meaningful motion and appearance
features, computes anomaly scores, and localizes suspicious regions using heatmaps and bounding boxes. A
real-time sound alert mechanism is integrated to notify operators immediately when abnormal crowd behaviour
is detected. Experimental evaluation on standard benchmark datasets demonstrates high detection accuracy,
reduced false positives, and strong real-time performance. The proposed approach significantly improves
surveillance automation, reduces manual dependency, and contributes to safer public environments.