System-Level Integration of Vision-Assisted Mission Triggering and Cascaded Control for Simulated Indoor Quadrotor Inspection
Abstract
This paper describes the implementation and indoor validation of an embedded-hardware quadrotor prototype for a simulated patrol task with a physical surrogate target. The platform integrates a TM4C123GH6PZI7 flight controller, an STM32 ground station, onboard YOLOv5s perception, MID360 LiDAR localization, downward time-of-flight altitude feedback, quaternion-based FusionAHRS attitude estimation, and a cascaded PID controller with actuator-saturation handling. The system separates periodic flight stabilization from asynchronous mission-event messaging and routes attitude, horizontal-state, and altitude observations through separate feedback channels. Five physical trials were performed in a 4.5 m $\times 3.5$ m indoor test area with a prescribed altitude setpoint of 1.20 m; the altitude variation observed across the five trials was approximately ±0.10 m. On a held-out single-class surrogate-target test split comprising 49 images and 374 annotated instances, the deployed YOLOv5s detector achieved an mAP@0.5:0.95 of 96.60%. Direct profiling on the MaixCam yielded total detector-processing times of 11.5–14.8 ms/frame under the measured conditions. In all five trials, the system completed one-touch takeoff, permitted-grid coverage, route display, target-statistics reporting, return, and landing. These findings demonstrate reproducible end-to-end operation within the reported indoor test envelope.