ETD

• active perception
• bias-aware visual servoing
• confidence-guided reframing
• dataset bias
• deep neural networks
• environmental monitoring
• gradient-based planning
• mobile robotics
• object detection
• scale bias
• spatial bias
• viewpoint planning

This thesis presents a bias-aware visual servoing pipeline for robotic environmental monitoring with DNN object detectors whose reliability depends on viewpoint. Detector confidence is analyzed offline to estimate two priors: an image-plane spatial map, and a scale-related profile. During operation, these priors are queried online (including their gradients) to generate reframing references that steer the robot toward viewpoints expected to yield more stable detections, without retraining the detector. A finite-state mission manager coordinates global coverage and local inspection.This strategy steers the robot toward viewpoints that are expected to be favorable for the detector, according to the extracted bias priors.The overall system is implemented and evaluated in ROS 2 simulation.