Computer vision and pursuit–evasion game theoretical controls for ground robots

Dan Shen,Haibin Ling,Erik Blasch,Khanh Pham,Genshe Chen

doi:10.1177/1687814019872911

Abstract

A hardware-in-loop control framework with robot dynamic models, pursuit–evasion game models, sensor and information solutions, and entity tracking algorithms is designed and developed to demonstrate discrete-time robotic pursuit–evasion games for real-world conditions. A parameter estimator is implemented to learn the unknown parameters in the robot dynamics. For visual tracking and fusion, several markers are designed and selected with the best balance of robot tracking accuracy and robustness. The target robots are detected after background modeling, and the robot poses are estimated from the local gradient patterns. Based on the robot dynamic model, a two-player discrete-time game model with limited action space and limited look-ahead horizons is created. The robot controls are based on the game-theoretic (mixed) Nash solutions. Supportive results are obtained from the robot control framework to enable future research of the robot applications in sensor fusion, target tracking and detection, and decision making.

Highlights

Pursuit–evasion games1–3 are mathematical tools to analyze the conflicting situations of two actors: a pursuer and an evader
Pursuit–evasion (PE) games are applied in the other areas such as geometry and graphs,7–10 sensor management,11,12 collision avoidance,13,14 multiple-player applications,15–19 general military operations,20 information deception,21 and high-level information fusion situation awareness
Motivated by the observation that PE games1–22 are mostly implemented in and tested by numerical simulations, we aim to develop a hard-in-loop robotic control framework to demonstrate various PE games and the associated data, sensor, and information fusion structures and solutions

Summary

Introduction

Pursuit–evasion games are mathematical tools to analyze the conflicting situations of two actors: a pursuer and an evader. The obtained video/images are processed by the visual tracking algorithms (see section ‘‘Visual tracking algorithms’’) and the estimated states are available to the Figure 1. The decision making follows the procedure: at the time k, the agent obtains the current robot states (Xp(k) for the pursuer and Xe(k) for the evader) from the visual tracking algorithm. It constructs and solves a discrete-time PE game. The overhead view camera captures an image at k + 1 and calls the visual tracker to estimate the robot states (Xp(k + 1) for the pursuer and Xe(k + 1) for the evader). The image-based robot position estimation can be used for robot control, enabling the PE game analysis

Background modeling and extraction of ROIs

Results

Conclusion