Vision-Centric Automated Parking Systems

• Vision-centric automated parking systems are platforms that use camera-based perception and deep learning to accurately detect parking slots and control vehicle maneuvers.
• They integrate both infrastructure- and vehicle-centric approaches, leveraging sensor fusion and semantic segmentation to enhance robustness in varied lighting and weather conditions.
• Key innovations include automated annotation, real-time trajectory planning with BEV integration, and digital twin connectivity for scalable smart parking solutions.

Vision-centric automated parking systems are software-physical platforms that utilize camera-based perception as the primary or sole means for detecting, mapping, localizing, planning, and controlling vehicle navigation and parking maneuvers in controlled environments such as parking lots and garages. These systems leverage monocular or multi-camera rigs (typically with fisheye lenses for near-field omnidirectional coverage), advanced image processing, deep convolutional neural networks (CNNs), geometric vision, and multi-sensor fusion to achieve robust, accurate, real-time occupancy detection, slot localization, trajectory planning, and end-to-end automation of valet or self-parking tasks.

1. System Architectures and Sensing Modalities

Vision-centric automated parking system architectures can be broadly categorized into two classes: infrastructure-centric (static overhead/lot cameras monitoring infrastructure) and vehicle-centric (on-board multi-camera rigs enabling AVP and self-parking).

Infrastructure-centric Approaches

These deploy fixed cameras (mono or stereo, typically with wide FOV) at strategic positions overlooking parking lots. Image streams are processed by a back-end server or edge computing node, which extracts parking slot occupancy and, optionally, guides users via mobile/web applications. Systems such as the Django-based CV pipeline (Chandrasekaran et al., 2022), the edge-enabled YOLOv11/NGSI-LD/digital-twin stack (Luz et al., 2 Feb 2026), and transformer object detection (Nguyen et al., 2024) exemplify this approach. Manual ROI initialization is a practical bottleneck, increasingly alleviated via automated spot localization exploiting vehicle detection and temporal tracking (Nguyen et al., 2024, Grbić et al., 2023).

Vehicle-centric, On-board Approaches

Production-grade automated parking systems employ 4–6 surround-view fisheye cameras for real-time, high-coverage perception. These are tightly integrated with inertial measurement units (IMU), wheel encoders, and (optionally) ultrasonics or radar. The SVCS (Surround-View Camera System) architecture supports real-time SLAM, BEV (Bird’s-Eye-View) fusion, semantic segmentation, and object detection for integrated mapping and control (Heimberger et al., 2021, Musabini et al., 2024, Abate et al., 2023, Sha et al., 2024, Tripathi et al., 2020).

Sensor fusion is typically realized through factor-graph back-ends, EKF-based odometry, or multi-task CNNs operating on calibrated BEV mosaics. Calibration encompasses intrinsic (focal length, distortion) and extrinsic (rigid transforms to vehicle frame) parameters, with robust parametrizations such as Kannala–Brandt for realistic fisheye modeling (Li et al., 2023).

2. Vision Perception Modules: Slot Detection, Occupancy Classification, and Semantics

Parking-Slot Detection

Traditional modules rely on explicit marking recognition—homographic projections, edge/Hough transforms, and geometric verification—operating on IPM-transformed images to generate slot hypotheses (Heimberger et al., 2021, Jo et al., 2021). Recent advances employ object detection backbones (YOLOv5, DETR, Polygon-YOLO) and geometric tracking (e.g., PakLoc, APSD-OC) to achieve fully automated, annotation-free slot localization via temporal consistency of detected vehicle boxes, clustering in BEV coordinates, and cluster filtering (Nguyen et al., 2024, Grbić et al., 2023).

Transformer-based cross-view networks now directly learn BEV slot and vehicle localization from several wide-angle physical views, correcting for perspective and lens distortion end-to-end (Musabini et al., 2024). Polygonal representations capture entrance lines, extent, and slot orientation up to 25m range, with localization errors of 20–25 cm (Musabini et al., 2024).

Occupancy Detection

Per-slot occupancy status is determined via classical thresholding/ROI statistics (Chandrasekaran et al., 2022), patch classification with deep CNNs (ResNet34, AlexNet) (Nyambal et al., 2021, Grbić et al., 2023), or one-stage detectors leveraging ROIs for object detection (PakSta, YOLOv11) (Nguyen et al., 2024, Luz et al., 2 Feb 2026). State-of-the-art systems now perform slot instance detection and direct occupancy labeling within a unified deep learning pipeline, eliminating dependency on grid-based manual labels and achieving AP75 >93% in varied weather/lighting (Nguyen et al., 2024).

Semantic Segmentation and Mapping

Semantic segmentation networks (U-Net, DDRNet, multi-task transformers) operate on BEV mosaics to robustly extract parking lines, arrows, guide signs, and other semantic entities critical for localization (Li et al., 2023, Qin et al., 2020, Musabini et al., 2024). Statistical feature extraction ensures slot landmarks serve as long-term stable anchors across illumination changes, occluded zones, and repeated textures (Qin et al., 2020, Sha et al., 2024). Slot entry edges and boundaries are encoded as semantic keypoints for mapping, registration, and functional localization (Sha et al., 2024).

3. Localization, Mapping, and State Estimation

Robust localization in GPS-denied, texture-poor environments—typical of indoor garages—relies on visual-inertial odometry (VIO), semantic mapping, and loop closure.

Visual-Inertial-Semantic SLAM

Surround-view systems implement tightly coupled SLAM pipelines integrating multi-camera input, IMU, wheel odometry, and semantic slot features within a factor-graph framework (Abate et al., 2023, Li et al., 2023, Sha et al., 2024). Key state variables encompass vehicle SE(3) pose, velocity, landmark depths, IMU biases, and explicit slot landmark positions. Back-ends optimize the joint graph with residuals from visual feature matches, odometry, IMU preintegration, and semantic slot association (Jiang et al., 2024). Specialized robustification (Max-Mixture, Cauchy kernels, semantic pre-qualification) enables rejection of incorrect slot associations and handles repetitive, aliased environmental structures (Huang et al., 2018, Li et al., 2023).

Semantic Anchoring and Slot Management

Explicit parking slot residuals serve as semantic anchors, dramatically reducing cumulative localization drift by periodically "pinning" the trajectory to a geometrically regular, repeatedly observed slot map (Jiang et al., 2024, Sha et al., 2024). Slot management modules include association via kd-trees, clustering, stability filtering, and adaptive weighting based on BEV distortion and confidence measures (Sha et al., 2024). These enable robust slot map construction, periodic slot filtering, and effective loop closure.

Free-space and Dense Mapping

Dense occupancy/freespace grids are constructed by CNN-based ground/obstacle segmentation, back-projected to BEV via homography, and aggregated to produce TSDFs or occupancy meshes (Abate et al., 2023). These facilitate safety-critical path planning, real-time obstacle avoidance, and high-fidelity environment modeling. BEV-augmented mapping is further robustified via flare removal and Fourier-based inpainting (Li et al., 2023).

4. Trajectory Planning, Control, and Decision Logic

Integrated planning-control stacks utilize BEV occupancy and slot map outputs to generate feasible, collision-free parking maneuvers.

Planning Strategies

Classical planners employ improved A* graph search with kinematic constraints (bicycle model), distance-weighted heuristics, bidirectional search, and trajectory smoothing (Bezier, B-spline) (Zhao, 2024). Numerical optimization (NLP) refines trajectory points to ensure kinematic feasibility, comfort, and safety. Tight integration with BEV grid updates supports real-time re-planning under dynamic scenarios.

End-to-End Learning for Parking Policy

Recent research implements imitation-learned policies mapping raw multi-camera input and goal tokens directly to control commands via transformer/ResNet backbones with BEV feature fusion (Chen et al., 14 Sep 2025). The Control-Aided Attention (CAA) mechanism trains the attention module with gradients from the control head, enforcing policy-relevant, stable attention in critical spatial zones—surpassing both conventional modular stacks and prior end-to-end baselines (TSR >87.5%) (Chen et al., 14 Sep 2025).

Decision Logic, Distributed Coordination

In distributed AVP, reservation managers orchestrate slot assignment and queuing for multi-vehicle fleets with vision-detected slot availability. Synchronization is achieved via message-oriented middleware (e.g., ROS2+Zenoh), ensuring deterministic, collision-free parking across distributed hosts (Islam et al., 22 Jan 2026).

5. Algorithmic Innovations, Robustness, and Scalability

Automated Annotation and Label Reduction

PakLoc and APSD-OC pipelines demonstrate automated spot localization via vehicle detection and cross-frame clustering, eliminating >94% of manual labeling and supporting rapid deployment/maintenance in dynamic camera placement settings (Nguyen et al., 2024, Grbić et al., 2023).

Handling Viewpoint, Weather, Occlusion

State-of-the-art slot detection and occupancy classification frameworks demonstrate high invariance to camera viewpoint (≥95% accuracy across nine distinct camera angles), weather class (accuracy/AP variation <5%), illumination (day/night, rain), and partial occlusion (recall drop limited to ≈83% in most challenging views) (Grbić et al., 2023, Nguyen et al., 2024).

Resource Efficiency and Edge Computing

Edge platforms such as Raspberry Pi 3B+ running YOLOv11m-TFLite with distance-aware spot matching and ABBP achieve balanced accuracy of 98.80% at 8 s inference per frame, enabling privacy-preserving, local occupancy processing, and IoT-based distributed deployment suitable for campus-scale and multi-lot applications (Luz et al., 2 Feb 2026).

Sensor Calibration and Failover

Periodic fisheye calibration (e.g., using the PEFT model in MT F-CVT) and soiling detection modules are essential for maintaining BEV transform integrity and overall system robustness (Maddu et al., 2019, Musabini et al., 2024). Redundancy via ultrasonics, radar, or fallback to alternative sensing modes is integrated for functional safety (Heimberger et al., 2021).

6. Performance Benchmarks and Limitations

Typical localization errors range from mean 2.36–5.23 cm in centering parking maneuvers for top semantic SLAM systems (Qin et al., 2020), trajectory drift under 1% of path length for multi-camera BEV fusion pipelines (Abate et al., 2023), and slot/vehicle detection F1-scores up to 0.89 with cross-view transformers (Musabini et al., 2024). Automated pipelines yield up to 99% detection/classification accuracy on challenging datasets including PKLot and CNRPark+EXT (Nyambal et al., 2021, Grbić et al., 2023, Nguyen et al., 2024).

Limitations include residual sensitivity to extreme lighting/glare (especially for template-based detection), degraded performance under heavy lens soiling, edge-case failure for highly dynamic/occluded scenarios, and the need for re-validation when vehicle/camera geometry changes (Jo et al., 2021, Sha et al., 2024, Musabini et al., 2024).

7. Extensions, Digital Twins, and Future Directions

Vision-centric systems integrate seamlessly with digital-twin/Smart City stacks via NGSI-LD, FIWARE, and IoT middleware for real-time dashboarding, AI-agent orchestration, system-wide analytics, and predictive simulation of campus/facility-level parking dynamics (Luz et al., 2 Feb 2026). Emerging directions include end-to-end differentiable BEV pipelines, geometric module unification, robust handling of ramps/multistory environments (with 6-DoF SLAM), and formal safety verification for deep vision modules (Sha et al., 2024, Heimberger et al., 2021).

In sum, vision-centric automated parking systems now constitute a mature technological discipline, featuring high-accuracy perception pipelines, robust localization and control architectures, automated annotation and digital integration, and validated performance across diverse environments, all built atop advanced geometric and deep learning methodologies.