Year: 2026
Type: dissertation
Abstract: UAV-based infrastructure inspection requires visual perception systems that balance detection reliability with limited onboard compute, energy, and latency budgets. A static detector deployment fixes this tradeoff: a lightweight model is efficient but may diverge from a stronger detector, while a larger model provides stronger reference behavior at higher computational cost. This thesis investigates DMS as a runtime mechanism for selecting between fast and accurate YOLO-family detectors during UAV inspection. The proposed framework evaluates DMS as a conditional systems tradeoff rather than a guaranteed speedup. It combines deterministic switching policies, deployed runtime accounting, accurate-reference agreement, random-switch baselines, trigger validity analysis, and CPU/CUDA platform comparison. The main benchmark contains 14 insulator videos from glass and porcelain domains, 84 completed runs, and 672 run-policy rows. Because dense frame-level human annotations are unavailable for these videos, the accurate detector is used as the reference model; the main metric is therefore accurate-reference agreement, not dense human-labelled video mAP. The results show that DMS can create intermediate runtime-quality operating points between fast-only and accurate-only deployment. Confidence EMA is the strongest runtime-aware canonical candidate because it avoids separate image-proxy computation and uses the fast detector as a watchdog. Local-contrast hysteresis provides strong exploratory agreement-oriented evidence, but is not treated as a universal winner. CPU results show useful intermediate tradeoffs, whereas CUDA results reveal platform dependence: when accurate-only inference is already fast, proxy overhead can dominate the saved detector runtime. Structural-texture fea- tures are predictively associated with accurate-model benefit, but no single scalar trigger is universally sufficient. Jetson Nano experiments further support the embedded runtime and energy motivation. The thesis concludes that DMS is useful only under measurable conditions: the accurate detector must add benefit over the fast detector, the trigger must outperform same-usage random switching, and deployed runtime or energy economy must remain favorable.
Author: Chirravuri, Ganapathi Teja
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Technische Universität Chemnitz
OpenAlex ID: W7225554571
