Extending Navigation Service under Sensor Failures: An Approach by Integrating System Identification and Vehicle Dynamic Model

Abstract

Localization plays a vital role in various autonomous systems, providing essential information for perception and planning tasks. However, mainstream localization methods are based on the sensors approach, which is vulnerable in some extreme conditions where sensors probably fail in a short period, such as the camera-based visual positioning. This study proposes a sensor-free localization method by integrating vehicle dynamic models and an online system identification module. First, a system identification process is conducted online to identify the system dynamics of the powertrain system and the steering system of the autonomous vehicle. Then, the identified system responses are taken as the control input of the vehicle dynamic model to produce the positioning results. The simulated experiments show that the proposed method achieves better positioning performance than the conventional vehicle dynamic models. In addition, the extendibility of the proposed method is explored by fusing it with extra sensors based on the extended Kalman filter (EKF). Furthermore, the navigation ability of the proposed method without sensors is also examined along a trajectory of 140 meters. The proposed method successfully accomplishes the navigation task without any collisions, demonstrating the effectiveness in enhancing the security of autonomous systems with navigation needs when sensors fail in extreme conditions.

Problem Addressed

Mainstream sensor-based localization methods are vulnerable in extreme conditions where sensors may fail for a short period, such as camera-based visual positioning, threatening the safety of autonomous systems that depend on continuous navigation.

Key Contributions

• Sensor-free localization integrating vehicle dynamic models with online system identification
• Online identification of powertrain and steering dynamics using ARX models recursively updated via RLS
• Identified system responses used as control inputs of the vehicle dynamic model for positioning
• EKF-based fusion with extra sensors demonstrating extendibility to mainstream positioning systems

Conclusions

In Gazebo simulation, the proposed VDM-OnSI method largely outperformed the conventional vehicle dynamic model with a more than 77% reduction in ATE, and EKF fusion with extra sensors further boosted accuracy. When sensors failed, VDM-OnSI was the only method that completed a 140-meter navigation path without collisions. Limitations include the empirically chosen ARX model order and the trade-off between model complexity and the computational cost of repeated identification.