Аннотация
This article presents an event-based edge video surveillance architecture for Internet of Things (IoT) systems, in which the Passive Infrared (PIR) sensor initiates frame capture, the ESP32-CAM module (OV2640) performs control and network publishing, and the Mojo Field-Programmable Gate Array (FPGA) performs hardware-accelerated preprocessing and detection. Communication between the ESP32-CAM and FPGA is carried out via the Universal Asynchronous Receiver–Transmitter (UART) interface at a rate of 921,600 baud (optionally using INT/READY signals), whereas detection results are transmitted to the cloud via Wi-Fi using Message Queuing Telemetry Transport (MQTT) or Hypertext Transfer Protocol (HTTP). Two modes of operation are experimentally evaluated: ESP32-CAM only and ESP32-CAM + FPGA. In terms of end-to-end latency (from PIR triggering to acknowledgment received), a reduction in latency quantiles is observed. For N = 80 events, the P50 decreases from 620 to 480 ms, the P95 from 980 to 780 ms, and the P99 from 1,300 to 1,050 ms, confirming the performance gains achieved by offloading computation to dedicated hardware. The UART load is quantitatively characterized based on the transmitted data type. For compressed frames, the average packet size is approximately 1,200 B (P95: 1,600 B) at a rate of 25 packets/s, corresponding to approximately 30,000 B/s. For Regions of Interest (ROIs) or compact features, the average packet size is approximately 180 B (P95: 240 B) at 40 packets/s, corresponding to approximately 7,200 B/s. For detection results (bounding boxes, confidence values, and flags), the average packet size is approximately 64 B (P95: 96 B) at 40 packets/s, corresponding to approximately 2,560 B/s. These results demonstrate the advantage of transmitting feature-level data instead of full-frame data in bandwidth-constrained scenarios. Detection performance is evaluated using annotated views (60 windows per view). For human movement at a distance of 2 m (front view), accuracy/recall/F1-score values of 0.93/0.95/0.94 are achieved, whereas at 4 m (side view), the values are 0.87/0.85/0.86. The False Alarm Rate (FAR) is 0.15 and 0.25, respectively. For scenes without target movement, the FAR ranges from 0.03 to 0.17, depending on background conditions (idle scene, background movement, and lighting changes). The results demonstrate that the combination of the PIR sensor, ESP32-CAM, and FPGA provides an effective trade-off between latency, communication overhead, and detection performance, making it suitable as a minimal yet extensible platform for distributed security systems and industrial event-based monitoring.