A Cost-Effective Autonomous Parking System Using a Two-Channel Vision Sensor and FSM

초록

High-precision autonomous parking systems typically rely on expensive sensor suites. These include LiDAR, radar, and GPS. However, integrating multiple sensors significantly raises both costs and technical complexity. To overcome this, this paper presents a cost-effective, vision-based autonomous parking system. It uses only two 120-degree wide-angle cameras. For parking spot recognition, a deep learning model was developed based on YOLOv8. This model works with a Bird's Eye View transformation. This determines the relative positions of the vehicle and the parking spot. The system's core control is a finite state machine-based reactive model. This approach offers predictable, deterministic operation sequences. It helps mitigate reliability concerns often seen with vision-only systems. In 300 simulation tests, the system demonstrated practical performance in the Gazebo environment. The average "pure"parking time was 30.11 seconds. The mean position error was 0.1505 m, and the mean angular error was 3.22 °. 90.33% parking success rate was also achieved. This research confirms that practical, high-efficiency parking is possible without expensive sensor fusion. This study is expected to facilitate progress in the practical development of vision-based autonomous parking technology.

키워드