A Comparative Study on Autonomous Vehicle Local Path Planning Through Model Predictive Control and Frenet Frame Method

In recent years, autonomous vehicles (AVs) have been receiving increasing attention from investors, automakers, and academia due to the envisioned potentials of AVs in enhancing safety, reducing emissions, and improving comfort. The crucial task in AV development boils down to perception and navigation. The research is underway, in both academia and industry, to improve AV's perception and navigation and reduce the underlying computation and costs. This article proposes a model predictive control (MPC)-based local path-planning method in the Cartesian framework to overcome the long computation time and lack of smoothness of the Frenet method. A new equation is proposed in the MPC cost function to improve the safety in path planning. In this regard, an AV is built based on a 2015 Nissan Leaf S by modifying the drive-by-wire function and installing environment perception sensors and computation units. The custom-made AV then collected data in Norman, Oklahoma, and assisted in the performance evaluation of the two control algorithms in this work. Both straight roads and curved roads are considered in the evaluation. For the purpose of saving costs and raising real-world implementation potential, the vision-only solution is applied in object detection and bird's-eye-view coordinate data generation. MPC and Frenet coordinate system approaches are independently employed to generate a safe and smooth path for the AV using the collected data. The two methods are compared in terms of smoothness, safety, and computation time. Compared with the Frenet-based method, the proposed MPC method reduces the computation time by 80%, and the path smoothness is significantly improved.