Automated surveillance is essential for the protection of Critical Infrastructures (CIs) in future Smart Cities. The dynamic environments and bandwidth requirements demand systems that adapt themselves to react when events of interest occur. We present a reconfigurable Cyber Physical System for the protection of CIs using distributed cloud-edge smart video surveillance. Our local edge nodes perform people detection via Deep Learning. Processing is embedded in high performance SoCs (System-on-Chip) achieving real-time performance (≈ 100 fps - frames per second) which enables efficiently managing video streams of more cameras source at lower frame rate. Cloud server gathers results from nodes to carry out biometric facial identification, tracking, and perimeter monitoring. A Quality and Resource Management module monitors data bandwidth and triggers reconfiguration adapting the transmitted video resolution. This also enables a flexible use of the network by multiple cameras while maintaining the accuracy of biometric identification. A real-world example shows a reduction of ≈ 75% bandwidth use with respect to the no-reconfiguration scenario.