Research on dynamic obstacle avoidance and complex path planning strategies based on ROS robots

Yiwei Chen, Jr-Hung Guo, Kuo-Hsien HsiaProceedings of International Conference on Artificial Life and Robotics 2024 · DOI

Abstract

Robot Operating System (ROS) is a software system framework used by many robot systems. Although ROS provides a good development environment and related frameworks, ROS is not suitable for public places such as restaurants because of the coming and going of people. Dynamic obstacle avoidance is often handled by stopping the robot, or when there are frequent environmental map changes or when sensors such as optical radar fail, the stop action is also used. However, this often causes path obstructions or delays in completing tasks. Therefore, this study attempts to use images, auxiliary sensors, and various path avoidance strategies to solve the problem of the robot stopping and waiting for the obstacles to disappear and the problem of rapid changes in map paths.

Introduction

While the Robot Operating System (ROS) is the industry standard for development, robots often struggle in "social" spaces where people are constantly moving. Most current mobile robots rely heavily on 2D LiDAR, which only scans a single horizontal plane. If an obstacle is above or below that laser line, the robot is effectively blind to it.

Furthermore, when faced with a moving obstacle, many ROS configurations default to a "stop and wait" behavior. This leads to inefficient path planning, task delays, and robots that feel "clunky" in real-world applications. This research aims to bridge this gap by using computer vision and more powerful onboard processing to enable smoother, more proactive obstacle avoidance.

System Architecture and Methods

The researchers moved away from low-power microcontrollers (like the Raspberry Pi) in favor of an x86-based computer (Intel N-series) to provide the necessary "horsepower" for real-time image processing.

The system architecture integrates an x86 PC running Ubuntu and ROS with various sensors including Camera and LiDAR.

Their software stack integrates three key algorithms to balance performance and reliability:

  1. KCF (Kernelized Correlation Filter): Used for high-speed initial target identification.
  2. DSST (Discriminative Scale Space Tracker): An improvement over KCF that handles changes in the size (scale) of an object, allowing for better distance estimation.
  3. TLD (Tracking-Learning-Detection): A sophisticated algorithm that can "re-learn" a target if it disappears behind an obstacle (occlusion) or rotates.

To optimize CPU usage, the robot employs a tiered strategy. It uses the lightweight KCF+DSST combo for general tracking. Only when an object is determined to be dangerously close does the system trigger the TLD algorithm to ensure the robot doesn't lose track of the obstacle during complex maneuvers.

Experimental Evaluation

The system was evaluated using both simulation and physical hardware:

  • Gazebo Simulation: A standard ROS tool used to generate dynamic environments with multiple moving objects of varying sizes and positions.
  • Real-world Hardware Testing: The algorithms were deployed on a physical mobile robot platform equipped with cameras, LiDAR, and an STM32 control board to verify performance in real-world conditions.

Key Results and Insights

Gazebo simulation results comparing standard KCF tracking with the enhanced KCF+DSST approach.
  • Efficiency Gains: The KCF+DSST tracking combination was found to be approximately 20% faster than using the TLD algorithm alone.
  • Robustness: While KCF+DSST struggled with "interlaced" objects (objects crossing paths) or rapid rotations, the integrated TLD system handled these scenarios correctly, maintaining a lock on the obstacle.
  • Reduced Downtime: By combining vision-based tracking with proactive path re-planning, the researchers reduced robot "stop-and-wait" instances by 90% compared to traditional LiDAR-only methods. The study found that 90% of robot pauses in dynamic environments are unnecessary; the remaining 10% occur only when a path is physically, completely blocked.
  • Scaling Limits: The system performed smoothly with up to 150 dynamic objects in the environment. Significant system pauses only began to occur when the object count reached roughly 300, marking the upper limit of the current hardware configuration.

Conclusion

The primary hurdle for real-world robot deployment in human spaces is the "planar blindness" of 2D LiDAR. This study demonstrates that vision is a requirement for effective navigation. By using computational tiering—running heavy "Learning" and "Detection" phases only when necessary—robots can maintain high frame rates without sacrificing reliability. Future work will focus on further optimizing these algorithms to handle "dynamic emergence" (sudden appearances of obstacles) to eliminate the final 10% of unnecessary stops.

Summary updated

Abstract

Robot Operating System (ROS) is a software system framework used by many robot systems.Although ROS provides a good development environment and related frameworks, ROS is not suitable for public places such as restaurants because of the coming and going of people.Dynamic obstacle avoidance is often handled by stopping the robot, or when there are frequent environmental map changes or when sensors such as optical radar fail, the stop action is also used.However, this often causes path obstructions or delays in completing tasks.Therefore, this study attempts to use images, auxiliary sensors, and various path avoidance strategies to solve the problem of the robot stopping and waiting for the obstacles to disappear.The problem of rapid changes in map paths.
BibTeX
@inproceedings{chen2024research,
  author = {Yiwei Chen and Jr-Hung Guo and Kuo-Hsien Hsia},
  title = {{Research on dynamic obstacle avoidance and complex path planning strategies based on ROS robots}},
  booktitle = {Proceedings of International Conference on Artificial Life and Robotics},
  year = {2024},
  doi = {10.5954/icarob.2024.os5-1},
}
Figure