In smart cities, video surveillance is essential for public safety, evolving beyond simple camera installations and centralized monitoring due to the overwhelming amount of footage that challenges human operators. To enhance anomaly detection, experts have developed sophisticated computer vision techniques that classify events as normal or abnormal. This book explores an end-to-end urban video surveillance system, which aims to address asymmetric threats through three key strategies: firstly, it employs a corrective signal called 'task-specific QoE' that considers contextual factors; secondly, it utilizes machine learning driven predictive systems and a method known as "similarity-based meta-reinforcement learning" for effective anomaly detection; and thirdly, it advocates for "zero-touch" self-management systems based on autonomous computing. This holistic approach ensures rapid adaptation and situational awareness, effectively meeting the demands of modern businesses and enhancing overall safety in dynamic urban environments.