Smart Parking Using Computer Vision and Edge Computing

Ziad Tamim / September 14, 2024

AIEdge ComputingSmart CitiesParking ManagementComputer Vision

Urban areas globally face significant congestion, often worsened by inefficient parking management. Traditional sensor-based systems, while effective, are costly and impractical for large-scale implementation due to the need for sensors in every parking slot. My project aimed to address these inefficiencies using modern, cost-effective AI solutions.

What was the problem?

Typical parking management solutions rely heavily on centralized systems requiring significant bandwidth to transmit data from numerous cameras to central servers. This approach is expensive, introduces security vulnerabilities, and is challenging to scale effectively.

My Solution

Edge Computing

I developed an edge-based smart parking system capable of real-time occupancy detection directly on-device. This drastically reduces bandwidth requirements, lowers costs, and enhances data security.

Custom Lightweight Model

To ensure efficient operation on resource-limited edge devices, I crafted a custom neural network based on MobileNet's depthwise separable convolution blocks:

Depthwise Convolution: Processes each input channel separately, reducing computational load.
Pointwise Convolution: Merges channel outputs efficiently with a 1x1 convolution.
Optimized Architecture: My custom model has only 16 layers compared to MobileNetV1's 28, tailored specifically for binary classification (occupied vs. empty parking slots).

Model Compression Techniques

To further enhance performance and reduce resource demands, I employed:

Knowledge Distillation

Knowledge Distillation: A "teacher-student" approach where a smaller custom model (student) learns from a larger, powerful ResNet50 model (teacher). This maintains accuracy while significantly speeding up inference.

Quantization

Quantization: Applied post-training quantization to reduce model precision from 32-bit floats to 8-bit integers, significantly decreasing the model size by 92% (from 325 KB to just 26 KB).

How the Smart Parking System Works

System Design Here’s how my edge-based parking occupancy system processes and classifies parking slots in real-time:

a) Live Video Frame Capture

The system starts by capturing video frames using a camera attached to a Raspberry Pi 5.
These frames show the full view of the parking lot.

b) Annotated Parking Slots (LabelMe)

Before processing, each parking slot is manually annotated using LabelMe software.
This provides polygon coordinates for every slot so the system knows exactly where to look in each frame.

c) Cropping the Parking Slot

Using the annotated coordinates, the system crops each slot from the full video frame using a custom function.
This isolates the area of interest — the individual parking slot.

d) Perspective Transformation

A perspective correction is applied to the cropped slot to make sure it appears as a standardized square, even if the camera angle is tilted.
This ensures accurate classification regardless of viewing angle or distortion.

e) Occupancy Classification (QCustom Model)

The cropped and transformed image is then fed into the QCustom model running directly on the Raspberry Pi.
The model classifies the slot as either:
- 1 → Occupied
- 0 → Empty
The classification history is stored in a buffer (deque) to smooth out predictions and avoid flickering results.

f) Sending to Central System & Displaying on UI

The final status of each slot is:
- Visualized locally on the Pi (only for testing performance)
- Sent to the centralised server in a compact JSON format
A Flask web server then displays the parking availability in real-time through a web app dashboard, helping users easily find available spaces.

Datasets and Evaluation

I tested my solution extensively across multiple datasets:

CNRPark-EXT Dataset: 144,965 images across varying conditions, achieving 98.69% accuracy.
Prototype Dataset: Controlled environment using toy cars, achieving 100% accuracy.
Hyderabad Dataset: Real-world dataset focusing on nighttime conditions, achieving 98.21% accuracy.

My compressed model ("QCustom") notably outperformed traditional heavyweight models in terms of speed (0.77ms inference time) and efficiency (only 26KB file size).

Results

The edge-based occupancy detection system successfully:

Reduced inference time by over 99% compared to MobileNetV1 (0.74ms vs. 156.5ms).
Achieved real-time performance at 30fps on Raspberry Pi 5.
Delivered 97.44% accuracy in real-world tests using a video feed.
Maintained accuracy across all datasets, even under challenging lighting conditions.
Matched or outperformed state-of-the-art methods, without requiring GPUs.

Final Thoughts

This project not only solved practical challenges in parking management but also highlighted my expertise in edge computing, deep learning, model optimization, and deployment of lightweight AI models. The methods and tools I mastered—including TensorFlow Lite, MobileNet architectures, knowledge distillation, and quantization techniques—demonstrate practical skills valuable for smart city initiatives and AI-driven IoT solutions.