Contemporary multi-camera surveillance systems face challenges in maintaining consistent identity throughout camera changes, especially in blind spots.Previous approaches involve appearance-only re-identification techniques. Appearancebased identification exhibits weakness in the event of lighting conditions change and lack of logic about impossible transitions. Therefore, we introduce the SentinelAI framework that separates the Multi-Target Multi-Camera (MTMC) problem into single camera tracking and graph-based cross-camera route deduction. The key concept is the directed-graph formulation of camera connectivity with time constraint imposed by physical limitation of speed. The proposed approach allows us to prove that the mathematical representation reduces the cross-camera identification search from O(N 2C) to O(C log N) and always provides valid solutions. The threat assessment engine combines signals from loitering, zone incursion, movement anomaly detection and crowd density with the method of exponential decay. The system is validated through the simulation of a residential setting using Unity three-camera prototype.

multi-camera tracking,graph-based route deduction,person re-identification,blind-spot inference,edge AI,compound threat scoring