9th International Conference on ICT for Sustainable Development

Authors - Vishalsinh P Gohil, Chetankumar Chudasama
Abstract - In the field of surveillance video analysis, the identification of anomalous occurrences is of utmost importance in order to guarantee security and safety. Nevertheless, current techniques frequently encounter difficulties in precisely detecting abnormalities within intricate contexts and fluctuating environmental circumstances. In order to tackle these problems, this research suggests a novel strategy that combines two complementing techniques: Multiple Segment Learning (MSL) ranking loss and a fusion of object identification and posture estimation. The suggested fusion technique synergistically integrates the advantages of both methodologies to optimise the precision and effectiveness of abnormal event detection. The MSL ranking loss model is trained for robust anomaly detection by dividing surveillance films into defined segments during training and employing both positive and negative containers. Simultaneously, YOLO is utilised for object recognition while pose estimation methods are employed to extract information about object classes and human-related poses. The integration of these streams not only boosts the accuracy of anomaly detection but also improves the identification of events that are not connected to humans. This makes the suggested technique adaptable and efficient in many surveillance settings.