Spot-Wise Smart Parking: An Edge-Enabled Architecture with YOLOv11 and Digital Twin Integration

Smart parking systems help reduce congestion and minimize users’ search time, thereby contributing to smart city adoption and enhancing urban mobility. In previous works, we presented a system developed on a university campus to monitor parking availability by estimating the number of free spaces from vehicle counts within a region of interest. Although this approach achieved good accuracy, it restricted the system’s ability to provide spot-level insights and support more advanced applications. To overcome this limitation, we extend the system with a spot-wise monitoring strategy based on a distance-aware matching method with spatial tolerance, enhanced through an Adaptive Bounding Box Partitioning method for challenging spaces. The proposed approach achieves a balanced accuracy of 98.80% while maintaining an inference time of 8 seconds on a resource-constrained edge device, enhancing the capabilities of YOLOv11m, a model that has a size of 40.5 MB. In addition, two new components were introduced: (i) a Digital Shadow that visually represents parking lot entities as a base to evolve to a full Digital Twin, and (ii) an application support server based on a repurposed TV box. The latter not only enables scalable communication among cloud services, the parking totem, and a bot that provides detailed spot occupancy statistics, but also promotes hardware reuse as a step towards greater sustainability.

💡 Research Summary

The paper presents a comprehensive edge‑enabled smart parking system deployed on the Institute of Computing building (IC‑2) at the University of Campinas (UNICAMP). Building on earlier work that estimated overall parking availability by counting vehicles within a region of interest (ROI), the authors advance to a spot‑wise monitoring approach that determines the occupancy status of each individual parking space. The core of the system is a lightweight YOLOv11m model, converted to TensorFlow‑Lite (TFLite) format, occupying only 40.5 MB. This model runs on a Raspberry Pi 4B, achieving an inference time of roughly 8 seconds per frame—acceptable for near‑real‑time operation on a resource‑constrained edge device.

To assign detected vehicles to specific spots, the authors replace the traditional mask‑overlap method with a distance‑aware matching algorithm. For every detection, the Euclidean distance between the detection’s centroid and the centroids of all predefined parking spots is computed; the spot with the smallest distance within a configurable threshold is marked as occupied. This spatial tolerance makes the system robust against minor camera misalignments, lighting variations, and slight vehicle position shifts that would otherwise cause false positives or missed detections.

Challenging parking zones—where lane markings are faint or the geometry is irregular—are handled through Adaptive Bounding Box Partitioning. The algorithm dynamically partitions the image into sub‑ROIs that better fit each spot’s shape, allowing the detector to focus on appropriately sized regions and improving accuracy in problematic areas.

Beyond detection, the paper introduces two architectural innovations. First, a “Digital Shadow” is created using NGSI‑LD compliant smart data models, providing a virtual representation of the parking lot that can be visualized and queried by other smart‑city services. Although currently a one‑way mirror of the physical lot, this component establishes a foundation for a full Digital Twin, which would enable bidirectional, real‑time data exchange and potentially control actions in future iterations.

Second, the authors repurpose a commercial TV‑Box (an Android‑based set‑top box) as an application support server. This device hosts MQTT brokers, a lightweight database, and a chatbot interface that delivers spot‑level occupancy statistics to users. By offloading communication, data aggregation, and user‑facing services to the TV‑Box, the system reduces reliance on external cloud infrastructure, lowers network latency, preserves privacy (image data never leaves the edge), and promotes circular‑economy principles through hardware reuse.

Performance evaluation over six months, comprising more than 12 000 annotated images, shows a balanced accuracy of 98.80 % (Precision ≈ 99.1 %, Recall ≈ 98.5 %). This outperforms prior campus implementations based on YOLOv5 or vehicle‑count ROI methods, which achieved 94.3 % and 91.7 % respectively. The authors also present statistical analyses of daily and monthly occupancy patterns, revealing peak usage periods and providing actionable insights for campus traffic management.

The paper acknowledges the remaining limitation of an 8‑second inference latency, which is still above true real‑time thresholds (<1 s). Future work includes model compression techniques (quantization, pruning), deployment on edge accelerators (NPU, Edge‑TPU), and extending the Digital Twin to support bidirectional control. Additional research directions involve multi‑camera integration, interoperability with smart‑traffic signals, and incorporation of electric‑vehicle charging data.

In summary, this work demonstrates a practical, low‑cost, and environmentally conscious smart parking solution that synergistically combines state‑of‑the‑art edge AI, robust spatial matching, adaptive region partitioning, and emerging Digital Twin concepts, thereby advancing both academic research and real‑world smart‑city deployments.