Screening Tunnel Configurations for Reduced Micro-UAV Hover Performance Using Marker-Based Visual Measurements

Micro-unmanned aerial vehicles (micro-UAVs) used for tunnel inspection must hover near structural boundaries that can affect the combined vehicle–environment response. This paper presents an outcome-based physical method for screening tunnel configurations through marker-linked flight measurements. A laptop-side convolutional neural network detects Quick Response (QR) markers in onboard video, and static calibration recovers a three-dimensional visual trajectory. Location labels assigned from the planned route associate each hover window with its target and boundary arrangement. Experiments in simplified channels characterize visual-to-Kalman measurement agreement before evaluation in a multi-configuration tunnel along TOP, MID, and BOT routes. Within command-derived hover windows, vision-to-target root mean square error measures placement relative to the commanded observation point, while visual position dispersion measures temporal fluctuation. Motion-capture-aided Kalman position and onboard attitude dispersion provide complementary measures. With the original distance mapping and static orientation model, the upper-and-lower-boundary groups have mean increases of 1.783 and 0.652 cm in target error and 2.047 and 0.577 cm in visual dispersion on TOP and BOT, respectively, relative to their open-side references. On MID, the group between opposing vertical surfaces has a target-error difference of 0.024 cm and a dispersion difference of −0.323 cm relative to the location opening toward the takeoff area. After recorded-attitude adjustment, BOT meets the screening rule under both the original and separate 6 cm distance mappings, whereas TOP meets it only under the original mapping. These results support route-specific assessment of hover performance within the tested layout and visit sequence.