Reinforcement Learning for Active Monitoring of Moving Equipment in 360-Degree Videos Academic Article uri icon

abstract

  • Computer vision techniques have been introduced recently to assist with visual surveillance of jobsite activities. However, multiple reality capture devices are needed to guarantee uninterrupted view of key objects. We propose to use deep learning with reinforcement learning (RL) to create a self-navigating active vision camera. The trained RL camera gains sufficient spatiotemporal knowledge to fix its gaze on an object by dynamically adjusting its position and view angle. We use a deep Q-learning network (DQN) to decide whether to move or rotate the camera to monitor moving forklifts in a 360-degree video. Results show that the RL camera can find a new position and angle with better view of the forklift in 73% of cases, and in the remaining 27% of cases, the visibility of the forklift remains unchanged. This indicates the effectiveness of the RL agent in locating the object of interest in complex and dynamic real-world settings.

published proceedings

  • Proceedings of the 29th EG-ICE International Workshop on Intelligent Computing in Engineering

author list (cited authors)

  • Nath, N. D., Cheng, C., & Behzadan, A. H.

citation count

  • 0

complete list of authors

  • Nath, ND||Cheng, C||Behzadan, AH