Trajectory Tracking Control Based on Visual Localization and Adaptive Radial Basis Function Sliding Mode Control

Autonomous driving is a prominent research focus in the field of intelligent vehicles, with trajectory tracking serving as a fundamental control strategy. However, most existing tracking methods rely heavily on GPS signals, making them susceptible to signal loss due to environmental obstacles such as buildings. Furthermore, traditional GPS-based approaches struggle to adapt to unpredictable real-world traffic conditions. To address these challenges, this study develops a vision-based localization model using a deep learning Convolutional Neural Network (CNN) to provide auxiliary localization signals. An Adaptive Radial Basis Function Sliding Mode Control (ARBFSMC) strategy is further employed to enhance the robustness of trajectory tracking. The proposed approach integrates the CNN-based localization model with the ARBFSMC controller using the CARLA simulator, leveraging real-time road images for training and testing. Comparative simulations between Pure Pursuit and ARBFSMC controllers were conducted. Results show that ARBFSMC achieves lower tracking errors than Pure Pursuit, especially at higher vehicle speeds. Under conditions where the CNN model’s predicted localization images remain consistent with the training set, the maximum average localization error is below 0.15 meters. Furthermore, when CNN localization outputs fluctuate, the ARBFSMC demonstrates greater resilience and bounded tracking behavior than Pure Pursuit, contributing to improved reliability in autonomous driving scenarios. Finally, real-world experiments were conducted on a GoKart platform utilizing RTK-GPS-based localization. When the vehicle was driven along a predefined trajectory at a constant speed of 5 km/h, the ARBFSMC controller achieved an average maximum tracking error of less than 0.37 meters. These results confirm the practical feasibility and bounded-control performance of the ARBFSMC strategy for real-world trajectory tracking applications.