Improved bounding box segmentation technique for crowd anomaly detection with optimal trained convolutional neural network.

Abstract

A crucial role in many security and surveillance applications is crowd anomaly detection, where seeing unusual activity helps avert possible threats or interruptions. For precise anomaly identification, current models might not successfully incorporate spatial and temporal features. To overcome these drawbacks, a novel Crowd Anomaly Detection based on Opposition Behavior Learning updated Chimp Optimization Algorithm (CAD-OBLChoA) is proposed in this research to enhance the detection of abnormal crowd behaviours in dynamic environments. In this research, bilateral filtering is used for smoothening the image and reducing noise for preprocessing phase. For object detection, a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM)-based bounding box approach is used. Then, features like Colour features, Shape features, and Improved Texture features are extracted. Finally, the anomalies get detected based on the trained extracted feature set in the system. For this, an optimized CNN is used, where training is done by the OBLChoA scheme via tuning the optimal weights. The proposed CAD-OBLChoA scheme achieved a higher specificity of about 0.924 and 0.931 in the 90% training data for datasets 1 and 2. This approach could significantly improve crowd monitoring and security, enabling faster identification of potential threats or emergencies.