Detect anomalies in a crowd using Deep Learning and Computer Vision

Abstract

This project stems from the Javaness R&D laboratory's initiative to enhance its expertise in deep learning algorithms for commercial innovation. Anomaly detection in crowds is a critical task for enhancing security, safety, and operational efficiency in public spaces, particularly during large-scale events like the Olympic Games 2024. Given the limitations of human ability to monitor live video surveillance effectively over extended periods, the development of artificial intelligence aims to improve attention and accuracy, reduce false alarms, and enable faster threat detection and intervention. This thesis presents the development and implementation of an end-to-end tool designed to detect and track objects and identify anomalies in human crowds using simulated video surveillance data. Leveraging advancements in deep learning and computer vision, the project encompasses multi-object detection, multi-object online tracking, and anomaly recognition. Key contributions include the training and evaluation of various state-of-the-art detection models on key performance metrics, benchmarking tracking algorithms for real-time application, and implementing effective baseline algorithms for anomaly detection across diverse scenarii. The system's performance was validated using comprehensive datasets, demonstrating its potential for real-time application. Additionally, a proof-of-concept web application was developed to showcase the practical implementation of the tool. While the project achieved promising results, it also identified areas for future improvement, including dataset diversity, inference times, pipeline optimization, and increasing the complexity of anomaly detection algorithms. These areas highlight the gap between the current prototype and a fully commercialized solution. Furthermore, the tool has been developed to be as versatile as possible, requiring enhancements and adaptations for real-world commercial application. This work lays a strong foundation for future research and practical implementations in crowd anomaly detection, emphasizing the importance of ethical considerations to respect individual freedoms and fundamental rights in order to avoid falling into authoritarian excesses such as mass surveillance.