Long-Term Tracking of Evasive Urban Target Based on Intention Inference and Deep Reinforcement Learning.

Area of Science:

• Robotics and Autonomous Systems
• Artificial Intelligence
• Computer Vision

Background:

• Urban target tracking by Unmanned Aerial Vehicles (UAVs) is crucial for public safety but challenged by evasive targets and complex environments.
• Existing methods struggle with target loss due to unpredictable movements and unstructured urban settings.

Purpose of the Study:

• To develop a robust hybrid target-tracking approach for UAVs in urban environments.
• To improve the long-term tracking of evasive targets by integrating intention inference and deep reinforcement learning.

Main Methods:

• A Convolutional Neural Network (CNN) based model infers target intentions by fusing environmental data and trajectory observations.
• A Deep Reinforcement Learning (DRL) framework develops a target search policy, modeled as a Deep Neural Network (DNN).
• The DRL policy is trained through interaction with the task environment to optimize search strategies.

Main Results:

• Target intention inference effectively guides UAV search operations, significantly enhancing target-tracking performance.
• The proposed DRL-based search policy demonstrates high robustness against uncertain target behaviors.
• Simulation results validate the improved effectiveness and reliability of the hybrid tracking approach.

Conclusions:

• The fusion of target intention inference and DRL offers a promising solution for challenging UAV target-tracking scenarios.
• This hybrid approach improves UAVs' ability to maintain track of evasive targets in dynamic urban landscapes.
• The method shows significant potential for enhancing public safety through more reliable surveillance and tracking capabilities.