Speaker
Description
Personnel Safety Systems (PPS) at laser-plasma accelerator facilities rely on card-access door controllers to protect radiation-controlled areas. A well-known weakness is tailgating: an unauthorized person follows an authorized entrant through the door before it closes, leaving no credential record that the PPS can act on.
This work presents a multimodal tailgating detection system developed for the CLAPA II laser accelerator at Peking University. Door-access events are obtained through an OPC UA server on the access-control network, while an overhead camera feeds a lightweight people-counting detector running on an edge device (NVIDIA Jetson Nano) on the control network. A fusion state machine aligns the two event streams within a door-open window and compares the authorized count from access logs with the visually observed pass count; a mismatch beyond hysteresis thresholds is raised as a tailgating alarm, corroborated over consecutive windows to suppress false positives. Conditional image recording stores frames only around valid access events, minimizing storage and bandwidth.
The fusion result is published to EPICS as standardized process variables following a facility:subsystem:function:attribute naming convention across Access, Vision, Fusion, and Interlock subsystems. Alarm severity fields follow the EPICS alarm model, and the interlock output is a fault-safe enum record wired to a hardwired relay: loss of health defaults to "block," requiring explicit human reset. The system is deployed across two network segments via an edge gateway.
We describe the architecture, PV design, and state-machine logic, and report the evaluation framework — detection rate, false-positive rate, and end-to-end latency — comparing access-only, vision-only, and fused approaches. Because all underlying components are standard (OPC UA, a commercial detector, and EPICS), the architecture serves as a reusable template for other EPICS-based accelerator facilities.
