Crowd Anomaly Detection Using Deep Learning

This paper introduces an image-based crowd counting method within the Audiovisual Crowd framework, designed to detect and warn against mass crowd stampedes or trampling caused by dangerous overcrowding. The approach utilizes deep convolution neural networks (CNNs) to analyze both visual and audio data, extracting meaningful features to identify anomalies in crowd behavior. The model is trained using CNN architectures and evaluated with Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics to ensure precise crowd estimation. Experimental results show that while the video-only approach effectively captures spatial information, it is more susceptible to challenges such as low-light and noisy environments. In contrast, the combined audiovisual model achieves enhanced robustness and accuracy, reaching an overall accuracy of 94% in detecting critical crowd anomalies.