PaperlayerView exact revision on arXiv

Source-linked AI summary

A Smart, Efficient, and Reliable Parking Surveillance System with Edge Artificial Intelligence on IoT Devices

Ruimin Ke, Yifan Zhuang, Ziyuan Pu, Yinhai Wang

arXiv:2001.00269v2eess.SP

TL;DR

Rapidly growing surveillance data make centralized cloud processing costly in bandwidth and latency for smart parking. The paper develops an edge-IoT parking surveillance system that distributes detection and server-side processing, achieving 95.6% detection accuracy over a three-month field deployment.

  • Problem

    Rapidly expanding Smart City surveillance data challenge cloud-based processing because it requires substantial bandwidth and can introduce transmission latency.

  • Method

    The system performs background-based and SSD detection on IoT edge devices, transmits compact detection results, and uses server-side tracking and occupancy judgment.

  • Results

    95.6% detection accuracy was achieved during a three-month real-world experiment in the Angle Lake parking garage.

  • Takeaways & Limitations

    The proposed system provides a smart parking surveillance approach designed for efficient online occupancy detection with reduced data transmission and reliable operation across environmental conditions.

Abstract

from arXiv · show

Cloud computing has been a main-stream computing service for years. Recently, with the rapid development in urbanization, massive video surveillance data are produced at an unprecedented speed. A traditional solution to deal with the big data would require a large amount of computing and storage resources. With the advances in Internet of things (IoT), artificial intelligence, and communication technologies, edge computing offers a new solution to the problem by processing the data partially or wholly on the edge of a surveillance system. In this study, we investigate the feasibility of using edge computing for smart parking surveillance tasks, which is a key component of Smart City. The system processing pipeline is carefully designed with the consideration of flexibility, online surveillance, data transmission, detection accuracy, and system reliability. It enables artificial intelligence at the edge by implementing an enhanced single shot multibox detector (SSD). A few more algorithms are developed on both the edge and the server targeting optimal system efficiency and accuracy. Thorough field tests were conducted in the Angle Lake parking garage for three months. The experimental results are promising that the final detection method achieves over 95% accuracy in real-world scenarios with high efficiency and reliability. The proposed smart parking surveillance system can be a solid foundation for future applications of intelligent transportation systems.

I. INTRODUCTION

The paper motivates edge-based smart parking surveillance by identifying cloud computing’s bandwidth and latency limits for rapidly growing surveillance data. It proposes an IoT-and-server architecture combining edge detection with server-side algorithms for efficient, accurate, and reliable online occupancy detection.

  • Cloud computing can consume substantial bandwidth and introduce transmission latency when processing rapidly expanding Smart City data.
  • Traffic surveillance generates major data transmission, processing, and storage workloads, while many existing systems focus on recording, offline analysis, or cloud computing.
  • Parking surveillance requires automatic and online occupancy detection to improve parking-space search efficiency, but it faces similar computing-workload challenges.
  • The system combines edge SSD-based detection with background-based detection, server-side tracking, and occupancy judgment for lighting and occlusion conditions.
  • The system is designed around smartness, efficiency, and reliability, defined respectively as automatic recognition, real-time online processing, and consistent performance across environmental conditions.
  • The proposed architecture splits computation between local IoT devices and servers, limits transmitted data, and targets real-time smart parking surveillance.

II. LITERATURE REVIEW

The literature review categorizes parking occupancy sensing into wireless sensor, moving sensor, and vision-based approaches. It positions the paper’s edge vision system as a response to vision-based data-transfer costs and the limitations of handcrafted or centralized processing.

  • Parking occupancy detection research is divided into wireless sensor network, moving sensor, and vision-based solutions.
  • Wireless sensor networks require multiple sensor nodes for multiple parking spaces and typically use efficient threshold-based or straightforward signal-processing methods.
  • Moving-sensor systems use phones, probe vehicles, or onboard sensors to collect parking information through crowdsensing, but their applicability remains limited to specific scenarios.
  • Vision-based systems can cover multiple parking spaces with one camera, reducing cost per space compared with approaches using one sensor per space or moving unit.
  • Vision-based surveillance can generate large data volumes that increase transmission cost and unreliability, motivating implementation on edge devices.
  • The paper proposes a vision-based solution with system-architecture and algorithm designs targeting smartness, efficiency, and reliability.

III. PROPOSED SOLUTION AND DESIGN

The proposed design distributes parking-surveillance processing across camera nodes, IoT devices, transmission modules, and a centralized server. Edge devices perform online detection and transmit compact results, while the server performs tracking and occupancy judgment.

  • A. Overview: The system comprises camera nodes, Raspberry Pi 3B IoT devices, cellular transmission modules, and a centralized server.
  • A. Overview: The architecture balances computational load, data-transmission volume, reliability, and scalability across system components.
  • B. Choice and Design of the Main Pipeline: Two edge computer-vision threads convert raw video frames into online detections, reducing transmitted data and enabling efficient updates.
  • B. Choice and Design of the Main Pipeline: The server runs modified SORT tracking and occupancy-judgment algorithms designed for occlusion and extreme lighting conditions.
  • B. Choice and Design of the Main Pipeline: The detection-based pipeline performs vehicle detection at the edge and sends bounding boxes to the server for matching, preserving flexible and scalable labeling.

1) The Classification-Based Pipeline and the Concern for Scalability

The classification-based pipeline requires parking-space labeling on IoT devices, creating maintenance and scalability concerns when camera views change. The detection-based design moves detection to the edge while retaining server-side matching and labeling, transmitting compact detection results instead of raw video.

  • Classification-based pipeline: Local labeling requires visiting each IoT device and repeating the process after camera-view changes, limiting flexibility and scalability.Camera changes include angle adjustments or zooming in or out.
  • Detection-based pipeline: Edge detection with server-side matching keeps parking-space labeling centralized and reduces relabeling labor when camera views change.Each edge device sends one frame to the server for labeling, rather than transmitting parking-space image patches continuously.
  • Edge detector: SSD-Mobilenet with TensorFlow Lite runs over 1 FPS on Raspberry Pi 3B, while transmitted bounding boxes, object types, and probabilities reduce data volume by thousands of times versus video.Mobilenet is selected for its lightweight structure and faster detection on computationally limited IoT devices.
  • Training data: MIO-TCD adds 110,000 traffic-surveillance camera frames covering nighttime, truncated vehicles, low resolution, and shadows to enhance SSD training.The paper reports this as the first use of MIO-TCD for parking detection.
  • Detection performance: Enhanced SSD produces fewer missed detections and no false detections in examples involving small, occluded, and snowy nighttime vehicles.The comparison is against SSD trained only on Pascal VOC.

2) Background-Based Detection at the Edge

Standalone SSD detection remains insufficient for universal real-world deployment because edge speed and training coverage limit performance in difficult conditions. The system therefore adds background-based detection, whose local efficiency and stability complement SSD in extreme scenarios.

  • Limitations: SSD running at 1 FPS does not meet real-time edge detection requirements, limiting use of video temporal information.The detector may also perform poorly because training data cannot cover all real-world scenarios.
  • Background-based detection: BG-based detection is added at the edge as a complement to SSD.It is followed by blob detection before downstream processing.
  • Background-based detection: BG-based detection operates in real time locally and is relatively more stable in extreme scenarios, although it performs less well under normal conditions.The method is sensitive to video noise and lacks classification ability.
  • Data transmission: Cellular connectivity is used because the field-test location lacks Wi-Fi or Ethernet, while Zigbee and LoRa are suited to shorter-distance communication.The system uses a 4G LTE modem connected to the Raspberry Pi.

E. Occupancy Judgement Pipeline and Algorithms

The server combines SSD-based and BG-based occupancy estimates to judge parking-space status under difficult lighting and occlusion conditions. SSD detections are matched to labeled spaces using a confidence-weighted overlap score, temporal thresholds, and conflict-resolution rules.

  • Occupancy judgement: The server computes SSD-based occupancy through matching and BG-based occupancy through multiple-object tracking before combining them.The combination targets extreme lighting and occlusion conditions.
  • SSD-based occupancy: The matching score V_ij multiplies intersection-over-union between space i and detection j by the square root of detection probability p_j.Only vehicle detections are retained, and the square root gives greater weight to overlap than the original probability.
  • Temporal filtering: Double thresholding uses previous occupancy status to filter invalid matching scores, with T_hmax greater than T_hmin.Occupied and vacant spaces use different thresholds based on their immediately preceding status.
  • Conflict resolution: When one detection matches multiple spaces, the space with the largest matching score is occupied and the others are vacant.If one space still has multiple detections, its status is occupied.

2) Modified SORT and BG-Based Occupancy Detection

The server combines background-based detection with tracking to infer parking occupancy and address missed detections caused by stopped vehicles, occlusion, and extreme lighting. A modified SORT rule reconnects objects using bounding-box overlap, while detection results are combined conditionally in difficult scenes.

  • Modified SORT: Object tracking uses edge background-model detections to eliminate false detections and generate vehicle tracks for server-side occupancy analysis.Only bounding-box location information is transmitted back to the server.
  • Modified SORT: Modified SORT reconnects newly assigned IDs to recently tracked objects when their bounding boxes have sufficient overlap.The rule uses IoU with an old object tracked within the past m seconds because deep re-identification features are unavailable.
  • BG-Based Occupancy Detection: Parking spaces are labeled vacant or occupied from track start and end locations when tracked time exceeds t_track seconds.Tracks moving from inside to outside indicate vacancy, while tracks moving from outside to inside indicate occupancy.
  • Final Detection: The SSD detector is primary in normal conditions, while background-based detection compensates for occlusion and extreme lighting cases.Background-based detection is more stable in extreme conditions, whereas enhanced SSD performs better under normal conditions.
  • Final Detection: Under an extreme-lighting warning, occupied SSD results remain occupied, while SSD-vacant spaces use background-based results.The warning is triggered from a metric comparing current background-based occupancy with current and previous SSD occupancy.
  • Modified SORT: The modified SORT algorithm resolves the ID switch caused when a vehicle stops to change direction, reducing parking-occupancy detection error.This failure is illustrated in Figure 4 for background-based detection with original SORT tracking.

IV. EXPERIMENTAL RESULTS AND ANALYSIS

The study used preliminary video and image data to set system parameters before field testing. Parameter values provide general parking-context references, but some settings require adjustment for specific camera geometries.

  • Preliminary Test and Parameter Setting: More than five hours of parking-lot video and over two thousand images were collected for preliminary testing and parameter setting.The data came from internet sources and the Angle Lake parking garage, with preliminary testing also conducted at STAR Lab.
  • Preliminary Test and Parameter Setting: Seven system parameters require setting, and the listed values in Table I serve as references for general parking contexts.Specific cases may require adjustments to optimize system performance.
  • Preliminary Test and Parameter Setting: Thmax is set larger than Thmin, with both thresholds increased when neighboring parking spaces overlap more from the camera angle.The adjustment reduces false matching under greater geometric overlap.
  • Preliminary Test and Parameter Setting: The occlusion percentage threshold should increase as overlap between neighboring parking spaces grows with the camera angle.This parameter is specifically adjustable to the installation geometry.

B. System Installation and Data Collection

The field deployment covered indoor and outdoor parking spaces using two IoT devices at the Angle Lake garage, with server validation frames transmitted every ten minutes. Over three months, accuracy varied mainly with outdoor lighting and indoor occlusion conditions.

  • System Installation and Data Collection: Two IoT devices monitored sixteen parking spaces on the third and sixth floors of the Angle Lake parking garage.The sixth floor represented outdoor parking, while the third floor represented indoor parking.
  • System Evaluation: The system achieved 95.6% detection accuracy over three months while transmitting one validation frame every ten minutes.The experiment categorized performance by weather, time, weekday or weekend, and indoor or outdoor floor.
  • Cloudy and Rainy Scenarios: 97.5% weekday and 99.2% weekend accuracy was achieved in cloudy outdoor conditions.Consistent lighting contributed to the strongest reported weather-specific performance.
  • Cloudy and Rainy Scenarios: 93.7% weekday and 96.2% weekend accuracy was achieved in rainy outdoor conditions, with raindrops occasionally blocking spaces.A camera shelter was used, but lens blockage remained the main cause of lower rainy-day accuracy than cloudy days.
  • Sunny and Foggy Scenarios: 85.7% weekday and 91.6% weekend accuracy was recorded in foggy outdoor conditions, with thicker fog associated with lower accuracy.Background-based detection was activated more often in sunny and foggy conditions to compensate for SSD performance.
  • Sunny and Foggy Scenarios: The third-floor indoor deployment had more consistent performance than the sixth-floor outdoor deployment across weather conditions.Outdoor errors mainly involved shadows, reflections, and fog, whereas indoor errors were more associated with occlusion.

3) Nighttime, Weekend, and Occlusion

The system reports stronger nighttime and weekend detection performance, while its pipeline addresses occlusion and extreme-lighting cases. Its one-frame-per-second processing supports frequent occupancy updates, although accuracy comparisons across systems are not directly valid.

  • Nighttime and Weekend: Nighttime detection accuracy exceeded daytime accuracy, and weekend performance exceeded weekday performance.
  • Occlusion and Extreme Lighting: Final results combine enhanced SSD detections with background-based detection when extreme lighting or occlusion warnings are triggered.Figure 6 identifies scenarios (c), (d), and (h) as cases where background-based detection was activated.
  • System Efficiency: 1 frame per second supports updating a parking lot’s occupancy status every one second in most cases.
  • Cross-System Comparison: Accuracy cannot be directly compared across systems because their inputs and designs differ and no widely accepted public parking video dataset exists.

1) Application in Smart Parking Management

Smart parking management uses occupancy information to improve parking-resource allocation, information dissemination, and space searching. It also encompasses prediction, dynamic pricing, and real-time guidance.

  • Smart parking management targets more efficient allocation of parking resources, dissemination of parking information, and parking-space searching.
  • The stated goals include high accuracy, robustness, and low cost.
  • Applications include parking prediction, dynamic parking pricing, and real-time parking guidance.

2) Application in Advanced Infrastructure Systems

Edge computing is presented as a component of future infrastructure management, while parking occupancy data can support connected vehicles and transportation-network functions. The evaluated system reports 95.6% overall accuracy across varied scenarios.

  • Advanced Infrastructure Systems: Edge computing is described as a critical component of future infrastructure management, especially with the emergence of 5G communication.
  • Advanced Infrastructure Systems: A cost-effective, real-time, reliable, and scalable parking system could support smarter, more efficient, and more sustainable buildings and roadways.
  • Connected and Automated Vehicles: Parking occupancy data can help automated vehicles find parking and support vehicle-to-infrastructure functions through parking facilities.
  • System Evaluation: 95.6% overall detection accuracy was achieved across indoor, outdoor, weather, occlusion, daytime, and nighttime scenarios.
Loading 2001.00269v2…