Active Robot Vision with Perception-Aware Navigation for Worker Monitoring

Ensuring reliable object detection is essential for vision-based safety monitoring on construction sites. Fixed-camera systems often encounter challenges from occlusions caused by obstacles and workers’ postures. While mobile robots can address these limitations through dynamic repositioning, the incorporation of their visual detection range into navigation strategies remains an open challenge, limiting detection consistency in complex environments. This study proposes an active monitoring navigation method based on a detection field, which enables robots to dynamically adjust their positions to keep targets within the detectable range of the vision sensor. Using deep reinforcement learning, the navigation policy optimizes the robot's movements around workers to maintain optimal visibility. This study also explores how different detection field configurations impact social distance in worker–robot interaction, while enhancing navigation and detection reliability. By optimizing the detection field, this approach advances active monitoring systems, promoting more reliable and safer worker–robot interactions in complex and dynamic construction environments.