Perception, Planning, and Control in Autonomous Parking: A Systematic Review from Bird’s-Eye View Reconstruction to Manoeuvre Execution

Engineering notes

Key topics: autonomous driving, bev, perception, planning, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous parking is one of the most demanding manoeuvres a vehicle can be asked to perform. The available space is small, the margins for error are narrow, and the vehicle must achieve an exact final pose while respecting non-holonomic constraints. A persistent gap between detection and execution yields infeasible trajectories or terminal positioning errors. This systematic review examines how bird’s-eye view (BEV) reconstruction supports each link in the perception–planning–control chain, from slot detection through trajectory generation and manoeuvre execution, and identifies where those links remain weakest. No registered review protocol was used. Between 7 January and 10 February 2026, we retrieved 338 records through a single automated engine (OpenAlex) reaching multiple indexed venues, complemented by manual citation chasing; collection was semi-automated and screening was manual. Seventy-two studies passed eligibility screening, of which 45 provide primary, parking-specific evidence; the remaining 27 are prior surveys or generic methodological contributions retained as background rather than as primary evidence. We included studies addressing BEV reconstruction, slot detection, manoeuvre planning, or control with simulated or experimental validation, and excluded duplicates and works reporting no performance evaluation. The synthesis follows five axes: multi-camera homography-based BEV reconstruction, automatic parking-slot detection, geometric and kinematic modelling, planning in confined spaces, and nonlinear control with explicit constraint handling. Across the 45 primary studies, BEV-based methods are reported to improve geometric consistency and support reliable slot detection, and Hybrid A* combined with NMPC recurs as the most frequently reported route to dynamically feasible trajectories, although no study compares these approaches under identical conditions and no quantitative comparison across studies was performed here. The weakest point of the field is integration: perception and control are rarely coupled in a closed loop, and no standardised evaluation framework has yet gained wide acceptance. Heterogeneity in scenarios, metrics, and sensing configurations limits the strength of the evidence, and modular architectures show gaps in perception–planning–control coupling and reproducible transfer to platforms such as Gazebo. No formal risk-of-bias assessment was performed.

Links and sources

Comments