Computer Vision-Based Vehicle Allotment System using Perspective Mapping

This paper proposes a cost-effective, computer vision-based smart parking system that utilizes YOLOv8 for vehicle detection and inverse perspective mapping to merge multi-camera views into a simulated 3D environment for guiding users to vacant spots.

The Gist

Imagine walking into a massive, multi-story parking garage. It's dark, the layout is confusing, and you're driving in circles, wasting gas and time, just trying to find one empty spot. You might end up blocking traffic, frustrating other drivers, and stressing yourself out.

Now, imagine if that garage had a "super-eye" that could see every single car and every empty spot instantly, and then guide you straight to the best spot without you ever having to guess.

That is exactly what this paper proposes. The researchers from the National Institute of Technology Rourkela have built a Smart Parking System that uses computer vision (cameras and AI) instead of expensive sensors to solve this headache.

Here is the breakdown of how it works, using some everyday analogies:

1. The Problem: The "Expensive Sensor" Trap

Traditionally, to know if a parking spot is empty, garages install sensors (like little radar guns or pressure pads) in every single spot.

• The Analogy: Imagine trying to count how many people are in a stadium by putting a pressure-sensitive mat under every single seat. It would cost a fortune to buy the mats, and if one breaks, your count is wrong.
• The Solution: This team says, "Why buy a sensor for every spot? Let's just use the cameras we already have (like CCTV) and let a smart brain figure it out."

2. The Brain: YOLOv8 (The "Super-Speedy Detective")

To make the cameras "see," they use an AI model called YOLOv8 (You Only Look Once).

• The Analogy: Old detection systems are like a detective who looks at a photo, zooms in on the top left, then the top right, then the bottom, taking a long time to find the suspect.
• YOLOv8 is like a detective who glances at the entire photo in a single split-second and instantly shouts, "There's a car here! There's a pillar there! And that spot over there is empty!" It's incredibly fast and accurate.

3. The Magic Trick: Inverse Perspective Mapping (IPM)

This is the coolest part. Cameras usually take pictures from an angle (like looking down a hallway), which makes things look distorted. A spot far away looks tiny, and a spot close up looks huge.

• The Analogy: Imagine looking at a long table from one end. The plates at the far end look like tiny dots, and the ones near you look giant. If you tried to count them, you'd get confused.
• IPM is like a magical "flattening" tool. It takes that angled, distorted view and mathematically "flattens" it into a perfect bird's-eye view (like a Google Maps satellite image). Suddenly, every parking spot looks the same size, and it's easy to see which ones are actually empty.

4. The 3D Map: The "Digital Twin"

Once the system knows where the cars and pillars are, it doesn't just show a list of numbers. It builds a 3D model of the parking lot.

• The Analogy: Think of it like a video game map. The system takes the 2D photos from four different cameras, stitches them together, and builds a virtual 3D replica of the garage in the computer.
• It plots the cars and pillars as blocks in this 3D space. Any space between the blocks that isn't occupied is marked as an "Open Spot."

5. How It Finds the Empty Spots

The system looks at the 3D map and asks: "Where are the gaps?"

• The Analogy: Imagine a game of Tetris. The cars are the blocks that have landed. The system looks at the empty spaces between the blocks and says, "Aha! That gap is big enough for another car."
• It then calculates the distance between pillars or other cars to estimate exactly how many cars can fit in a row.

Why Is This Better?

1. Cheaper: You don't need to buy thousands of sensors. You just need cameras (which most garages already have).
2. Smarter: It can handle tricky lighting and different car sizes better than old sensors.
3. Faster: It guides you to the nearest spot instantly, reducing traffic jams inside the garage.

The Bottom Line

This research is like giving a parking garage a GPS and a personal valet that never gets tired. By using clever math (Inverse Perspective Mapping) and a super-fast AI detective (YOLOv8), they turn a confusing, dark parking lot into a clear, organized 3D map that tells drivers exactly where to go.

In the future, this could mean walking into a garage, seeing a green dot on your phone showing the exact spot, and driving straight there without ever circling around. It's a small step toward making our cities less congested and our lives a little less stressful.

Technical Summary

Here is a detailed technical summary of the paper "Computer Vision-Based Vehicle Allotment System using Perspective Mapping":

1. Problem Statement

The paper addresses the inefficiencies of traditional indoor parking systems, particularly in large, multi-story facilities. Key challenges include:

• High Costs: Conventional sensor-based systems (e.g., PIR, ultrasonic, LiDAR) are expensive to install and maintain across vast areas.
• Navigation Difficulties: Indoor environments often suffer from poor visibility and complex layouts, leading to driver confusion, traffic congestion within the lot, and prolonged search times for parking spots.
• Sensor Limitations: Fixed sensors often struggle with coverage gaps, integration complexities, and varying environmental conditions (lighting, occlusion).
• Need for Automation: There is a critical need for a cost-effective, scalable solution that automates parking space detection and guides drivers without heavy reliance on human intervention or expensive hardware.

2. Methodology

The proposed system utilizes a Computer Vision approach combined with Inverse Perspective Mapping (IPM) to create a 3D representation of the parking environment. The workflow is divided into three stages:

A. Data Collection & Simulation

• Instead of relying solely on real-world footage (which can be inconsistent), the authors generated a realistic 3D simulation of an indoor parking lot using Spline.AI.
• Video footage was captured from four distinct camera viewpoints within this simulation to cover various angles, lighting conditions, and obstructions (pillars, walls).
• The dataset consists of 150 images (scaled to 640x640) annotated to detect two primary classes: Vehicles and Pillars.

B. Object Detection (Model Architecture)

• Algorithm: The system employs YOLOv8 (You Only Look Once), a state-of-the-art object detection model known for its speed and accuracy.
• Training: The model was trained using the Roboflow tool for annotation (generating PASCAL VOC XML files). The dataset was split 75% for training and 25% for testing.
• Configuration: Training utilized a batch size of 8 and 50 epochs with an early stopping mechanism to prevent overfitting.
• Goal: To precisely localize bounding boxes for cars and pillars in real-time.

C. Spatial Mapping & 3D Reconstruction

• Inverse Perspective Mapping (IPM): The system converts 2D image coordinates into 3D spatial coordinates.
• Depth Calculation:
  • The centroid of each detected object's bounding box is calculated: $C(x, y) = (\frac{x_{max}-x_{min}}{2}, \frac{y_{max}-y_{min}}{2})$.
  • Euclidean distance ($D_{xy}$) is calculated from the camera origin (0,0) to the centroid.
  • Depth ($z$) is derived using the inverse of the distance: $z = \frac{1}{D_{xy}}$.
  • Logic: Objects closer to the camera yield a higher inverse value (positive depth), while distant objects yield lower values.
• 3D Visualization: Using Matplotlib, the system plots vehicles and pillars in a 3D Cartesian coordinate system.
• Space Extraction: The algorithm identifies vacant spaces by analyzing the gaps between detected pillars and vehicles. It calculates the distance between pillars along the X-axis and divides by the standard parking width to estimate available spots. Overlapping regions from different camera views are merged to create a unified map.

3. Key Contributions

• Cost-Effective Solution: Replaces expensive ultrasonic/LiDAR sensors with standard CCTV cameras and deep learning, significantly reducing infrastructure costs.
• Multi-Camera IPM Integration: Unlike previous works using single-camera views, this system stitches data from four camera angles to create a comprehensive 3D map, reducing blind spots.
• 3D Cartesian Simulation: The system does not just detect occupancy; it visualizes the parking lot as a 3D plot, allowing for dynamic calculation of empty spaces between structural elements (pillars).
• Optimized Model Selection: The study rigorously compares YOLOv5, YOLOv7, and YOLOv8, demonstrating the superiority of YOLOv8 for this specific application.
• Theoretical Analysis: The paper includes a mathematical proof (Lemma 1) arguing that while centroid-based inverse distance is effective, the system's reliance on bounding box area and centroid logic provides robust depth estimation for parking space allocation.

4. Results and Analysis

The authors evaluated the performance of YOLOv5, YOLOv7, and YOLOv8 using Precision-Recall curves:

Model	Overall Accuracy	Car Detection Accuracy	Pillar Detection Accuracy
YOLOv5	84.0%	97.3%	70.6%
YOLOv7	89.5%	94.7%	84.2%
YOLOv8	98.4%	98.6%	98.2%

• Performance: YOLOv8 significantly outperformed previous versions, achieving near-perfect accuracy in detecting both vehicles and pillars.
• Balance: YOLOv8 maintained a superior balance between precision (minimizing false positives) and recall, which is critical for avoiding misidentifying a pillar as a car or vice versa.
• Visualization: The system successfully generated 3D Cartesian plots that accurately represented the parking layout and highlighted vacant spaces.

5. Significance and Future Work

• Urban Impact: The system contributes to Smart City initiatives by reducing urban congestion, optimizing space utilization, and enhancing the driver experience through real-time navigation.
• Scalability: The approach is highly scalable as it relies on software algorithms rather than hardware proliferation.
• Future Directions:
  • Real-World Deployment: Testing the system in actual indoor parking facilities with varying lighting and weather conditions.
  • IoT Integration: Connecting the system with mobile apps for user navigation and IoT for real-time data analytics.
  • Algorithm Refinement: Further improving accuracy in highly occluded scenarios and expanding the system to handle larger, multi-level parking structures.

In summary, this paper presents a robust, low-cost, and highly accurate framework for smart parking management, leveraging the power of YOLOv8 and perspective mapping to transform how indoor parking spaces are monitored and utilized.