Deep Learning for Crowd Behavioral Analysis and Intelligent Management

Savulla, Yogesh (2025) Deep Learning for Crowd Behavioral Analysis and Intelligent Management. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (531kB) | Preview

Abstract

Anomaly detection in crowds is a key domain of research in the field of computer vision with applications to crowd safety, event surveillance and smart transportation. The conventional surveillance is based upon the manual observation that is inaccurate and not efficient in the implementation at large scale. Increased research into deep learning, specifically convolutional and attention-based networks, have made it possible to model video signals in terms of spatiotemporal dependencies automatically and make those models more robust in status as odd behaviour among the crowd.

Design A hybrid Convolution + Transformer-based video autoencoder that learns through a dataset of video sequences an understanding of normal crowd behaviour and can identify abnormality by reconstructing the input and analysing the perceptual error. The model can capture such short-term temporal dependencies by using sequences of five frames inPed1 and Ped2 UCSD Pedestrian Datasets to predict the following frame. Mean Squared Error (MSE), Mean Absolute Error (MAE), Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR) are used to measure reconstruction quality. They are fed into LightGBM as classifier features to dictate anomaly occurrence. It had a high reconstruction fidelity (MSE: 0.000282, SSIM: 0.9821, PSNR: 36.55 dB), and robust classification accuracy (accuracy: 77.84%, precision: 0.5845, recall: 0.6474, F1-score: 0.6144).

The results show how using deep feature extraction on the temporal dynamics combined with a learned decision boundary on the perceptual metrics to achieve strong annotation-saving performance-based anomaly detection. This method can find its way in real time application in detecting unattended objects, unlawful access, or strange movements. Advancement work in the future may involve the use of object tracking, adaptation of an objective within different camera environments, and multimodal sensor fusion to enhance detection capability further and flexibility.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Hamill, David UNSPECIFIED
Subjects:	Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning Q Science > Q Science (General) > Research > Research--Equipment and Supplies > Scientific apparatus and instruments > Physical instruments > Detectors > Remote sensing > Electronic surveillance > Video surveillance Z Bibliography. Library Science. Information Resources > ZA Information resources > Research > Research--Equipment and Supplies > Scientific apparatus and instruments > Physical instruments > Detectors > Remote sensing > Electronic surveillance > Video surveillance
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	24 Aug 2026 12:41
Last Modified:	24 Aug 2026 12:41
URI:	https://norma.ncirl.ie/id/eprint/9609

Actions (login required)

View Item