Real-Time Illegal Parking Detection System Based on Deep Learning

Xuemei Xie, Chenye Wang, Shu Chen, Guangming Shi, Zhifu Zhao

Abstract

This paper presents a robust, real-time system for detecting illegally parked vehicles by combining the Single Shot MultiBox Detector (SSD) with a custom tracking and motion analysis pipeline. By optimizing SSD's anchor box aspect ratios using K-means clustering on a vehicle-specific dataset, the system achieved a 99% detection accuracy at 25 frames per second. The approach demonstrates significant resilience to challenging environmental factors such as weather and lighting conditions compared to traditional background subtraction methods.

Introduction

Traditional illegal parking detection systems typically rely on background segmentation, which identifies "blobs" that deviate from a static background model. While computationally inexpensive, these methods are notoriously fragile; they struggle with sudden lighting changes (e.g., clouds or shadows), heavy rain, and crowded scenes where vehicles overlap. Furthermore, some methods rely on specific features like wheel detection, which fail if the wheels are occluded by other objects or the vehicle's body.

This research aims to bridge this gap by moving away from background subtraction toward a deep-learning-based object detection approach. By utilizing a neural network that can identify vehicles directly, the system remains effective regardless of environmental noise or pixel-level fluctuations in the background.

Methodology

The proposed system follows a four-stage pipeline designed for both accuracy and real-time performance:

  1. Vehicle Detection (Optimized SSD): The system uses SSD, an end-to-end deep neural network that predicts object categories and bounding boxes in a single pass. To optimize it for traffic monitoring, K-means clustering was performed on a custom dataset to determine the most common vehicle aspect ratios. Replacing standard VOC-based aspect ratios with custom ones (specifically 0.5 and 0.7) reduced background noise and improved detection accuracy.
  2. Tracking (Template Matching): Once a vehicle is detected, the system tracks it using normalized template matching on grayscale images. This allows the system to follow a vehicle's movement between high-intensity detection frames without the full overhead of the neural network.
  3. Periodic Re-detection: To prevent "tracking drift" and identify new arrivals, the full SSD detector runs every 25 frames. New detections are matched to existing tracks using Intersection-over-Union (IOU) metrics.
  4. Motion Analysis: The system monitors the distance a vehicle moves within a user-defined Region of Interest (ROI). If a vehicle remains stationary (moving less than a specific threshold) for more than 15 seconds, an illegal parking alarm is triggered.
Flowchart of the proposed illegal parking detection and tracking system.

Datasets and Training

A specialized dataset was collected to train and evaluate the system, focusing on various traffic conditions including sunny days, rainy weather, heavy shadows, and crowded scenes. Unlike general datasets like PASCAL VOC, this dataset specifically addresses the perspectives and aspect ratios found in overhead traffic surveillance footage, which is critical for the K-means anchor box optimization.

Key Results and Discussion

The integrated system achieved a 99% accuracy rate in detecting illegal parking events on test videos. It maintains a processing speed of 25 FPS on standard consumer hardware (Core i7-4790 CPU), meeting the requirements for real-time deployment.

Robustness to Environment

Because SSD learns the high-level features of a car—such as edges, shapes, and textures—rather than just looking for changes in pixels, the system remains highly accurate during rain or shifting shadows. These are scenarios that typically cause traditional background-subtraction models to fail due to false positives from moving shadows or water reflections.

Vehicle detection performance under rainy conditions, demonstrating system robustness.

Domain-Specific Anchor Tuning

A key insight of this work is that "off-the-shelf" deep learning models can be significantly improved by tuning hyperparameters to the specific geometry of target objects. Narrowing the aspect ratios of the SSD default boxes to match cars improved the detection mean Average Precision (mAP) from 75% to 77% while simultaneously reducing computational overhead.

Conclusion

The proposed hybrid pipeline successfully balances the heavy-duty detection capabilities of SSD with the speed of template matching. While the system is highly effective, future work could involve automating the Region of Interest (ROI) setup via lane detection and exploring more advanced deep-learning-based trackers to handle extreme scale changes or rotations in chaotic intersections.

Summary updated

BibTeX
@inproceedings{xie2017real,
  author = {Xuemei Xie and Chenye Wang and Shu Chen and Guangming Shi and Zhifu Zhao},
  title = {{Real-Time Illegal Parking Detection System Based on Deep Learning}},
  booktitle = {ICDLT},
  year = {2017},
  doi = {10.1145/3094243.3094261},
}
Figure