Enhanced Visual SLAM for Collision-free Driving with Lightweight Autonomous Cars

Zhihao Lin, Zhen Tian, Qi Zhang, Hanyang Zhuang, Jianglin Lan

Introduction

High-performance autonomous driving systems typically rely on a "heavy" hardware stack, utilizing power-hungry GPUs for deep learning and expensive, bulky sensors like LiDAR or Radar. While effective for full-sized vehicles, this setup is impractical for lightweight, small-scale robots with limited battery life and payload capacity. Furthermore, existing vision-based methods often struggle with perception noise—where inconsistent feature points lead to poor localization—and planning trade-offs, such as the oscillations seen in Artificial Potential Fields (APF) or the inefficient trajectories produced by Voronoi diagrams.

This paper introduces a vision-based navigation system designed specifically for lightweight autonomous cars, capable of running entirely on CPU-only hardware. By integrating an enhanced ORB-SLAM3 perception engine with a Control Lyapunov Function-Control Barrier Function (CLF-CBF) path planner, the proposed framework achieves stable, collision-free driving in complex indoor environments using only a single RGB-D camera.

Proposed Methodology

The framework is divided into two primary modules: an enhanced visual perception engine and a stable, safety-constrained path planner.

1. Enhanced Visual Perception

The system utilizes ORB-SLAM3 as its foundation for real-time mapping and localization. To address the challenge of perception noise in indoor environments—where repetitive textures or moving shadows can create outliers—the authors implemented an additional outlier-culling layer. This layer employs Lucas-Kanade (LK) optical flow and epipolar constraints to filter out unreliable visual features. By ensuring that only high-confidence points are used for pose estimation, the system maintains high localization accuracy without the need for GPU-accelerated deep learning.

2. Stable and Safe Path Planning: CLF-CBF-QP-SRP

The core innovation of the planning stage is the integration of several mathematical control techniques into a unified optimization problem:

  • Control Lyapunov Function (CLF): This function mathematically guarantees that the vehicle will reach its target state, ensuring global stability.
  • Control Barrier Function (CBF): This acts as a safety envelope, defining a "safe set" of states and ensuring the vehicle never enters a collision zone.
  • Quadratic Programming (QP): A fast optimization method that balances the competing goals of the CLF (reaching the goal) and the CBF (avoiding obstacles) in real-time.
  • Shape Reconstruction Process (SRP): Because CBFs are computationally simplest when calculated for circular boundaries, the authors developed the SRP. This process approximates irregular or polygonal obstacles using a series of bounding circles. This "engineering hack" allows the CPU to solve complex obstacle avoidance problems at high frequencies by simplifying distance-from-center calculations.

Finally, the system integrates the Timed Elastic Band (TEB) algorithm for local planning, allowing for immediate, short-range adjustments to the trajectory based on real-time sensor data.

Experimental Evaluation

The system was evaluated in a custom-built Gazebo simulation environment designed to mimic a challenging indoor scene. The environment featured a square room populated with nine square hollow tables and various decorative objects like TVs and murals. The simulated hardware consisted of a four-wheeled car equipped with an Intel RealSense D435i RGB-D camera.

Performance Benchmarks

The proposed method was compared against two industry-standard benchmarks: Artificial Potential Field (APF) and Voronoi Diagrams.

  • Stability: While the APF method exhibited severe oscillations when approaching obstacles—leading to unstable driving behavior—the CLF-CBF approach maintained smooth, continuous trajectories.
  • Efficiency: The Voronoi diagram method, while safe, produced significantly longer and less efficient paths. The proposed method generated more direct trajectories to the target.
  • Computational Efficiency: The average computational time for the global planner was approximately 2.0176 seconds. The authors note that this is remarkably low for trajectory optimization of this complexity on a CPU-only architecture.
  • Safety: Across all test cases, the total CBF value remained above zero, confirming that the vehicle successfully avoided all collisions.

Conclusion and Future Work

The research demonstrates that sophisticated autonomous navigation does not strictly require high-end GPUs or LiDAR sensors. By combining geometric SLAM with robust mathematical control functions, the authors provide a viable path for lightweight robots to navigate complex environments safely.

However, the study notes certain limitations. The current implementation was tested only in simulation and focuses on static obstacles. Future work will involve deploying the system on physical hardware to test its resilience against real-world sensor noise and lighting variations. Additionally, the authors plan to extend the CLF-CBF framework to handle dynamic obstacles, potentially integrating machine learning to help the system adapt to more unpredictable, moving environments.

Summary updated

Abstract

The paper presents a vision-based obstacle avoidance strategy for lightweight self-driving cars that can be run on a CPU-only device using a single RGB-D camera. The method consists of two steps: visual perception and path planning. The visual perception part uses ORBSLAM3 enhanced with optical flow to estimate the car's poses and extract rich texture information from the scene. In the path planning phase, we employ a method combining a control Lyapunov function and control barrier function in the form of quadratic program (CLF-CBF-QP) together with an obstacle shape reconstruction process (SRP) to plan safe and stable trajectories. To validate the performance and robustness of the proposed method, simulation experiments were conducted with a car in various complex indoor environments using the Gazebo simulation environment. Our method can effectively avoid obstacles in the scenes. The proposed algorithm outperforms benchmark algorithms in achieving more stable and shorter trajectories across multiple simulated scenes.
BibTeX
@misc{lin2024enhanced,
  author = {Zhihao Lin and Zhen Tian and Qi Zhang and Hanyang Zhuang and Jianglin Lan},
  title = {{Enhanced Visual SLAM for Collision-free Driving with Lightweight Autonomous Cars}},
  year = {2024},
  eprint = {2408.11582},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.11582},
}
Figure