1. Introduction

2. Literature Review

3. Limitations of Existing Research

4. Methodology

4.6. Model Training:

Leveraging the computational power of a P100 GPU on the Kaggle platform, YOLOv8 models were trained with a focus on achieving an optimal balance between training efficiency and model performance. This involved the selection of key hyperparameters:

• Epochs (50): A total of 50 epochs were chosen to ensure comprehensive exposure of the models to the training data. Validation performance was rigorously monitored throughout training to prevent overfitting and determine the most suitable number of epochs.
• Image Size (640x640): A standardized image size of 640x640 pixels was selected for all input images. This choice balances the ability to capture essential image information with maintaining computational efficiency during training.
• Training Batch Size (16): The training batch size was set to 16 images per iteration, considering both GPU memory utilization and the benefits of gradient updates.
• Nominal Batch Size (64): Specific to the YOLOv8 architecture, a nominal batch size of 64 was employed to leverage gradient accumulation across multiple training batches, enhancing training stability and convergence.

the rigrous hyperparameter tuning process underscores the commitment to achieving efficient traing that fostes superior model performance on target tasks within our smart parking application, while also promoting model generalizability for real-world scenarios

Figure 3. Confusion Matrix for YOLOv8s for Vehicle Detection.

Figure 3. Confusion Matrix for YOLOv8s for Vehicle Detection.

Figure 4. Confusion Matrix for YOLOv8m for License Plate Detection.

Figure 4. Confusion Matrix for YOLOv8m for License Plate Detection.

Figure 5. Training Insights of Precision (Equation (4)) and mAP50-95 (Equation (7)) for Top 2 combinations according to TOPSIS analysis (Table 2).

Figure 5. Training Insights of Precision (Equation (4)) and mAP50-95 (Equation (7)) for Top 2 combinations according to TOPSIS analysis (Table 2).

Figure 6. System flow of the proposed design.

Figure 6. System flow of the proposed design.

5. Details of Approach

5.1. Extraction of Region of Interest:

A simple functionality of the OpenCV [17] library was created to define and export regions of interest from the full frames. The module allows the user to select an area of interest on the video by using the mouse. After the left mouse button is clicked, the module captures the coordinates of the mouse for the top-left corner of the selected area. When the left mouse button is released, the coordinates for the bottom-right corner are captured. Then, the module draws a rectangle for the defined region on the video. User can save the selected region by pressing any key function. This rudimentary functionality allows the user to capture desired portions of the video for further analysis.

Figure 7. Algorithm 1 enables users to define a region of interest (ROI) by dragging the cursor from point A to point B (or vice versa).

Figure 7. Algorithm 1 enables users to define a region of interest (ROI) by dragging the cursor from point A to point B (or vice versa).

Algorithm 1 ROI Extraction using OpenCV

1: Step1: Import necessary libraries    2: - OpenCV    3: Step 2: Define global variables:    4: - frame: current video frame    5: - roi_start: starting coordinates of the ROI    6: - roi_end: ending coordinates of the ROI   7: Step 3: Initialize the video capture object to read frames from the video source.   8: Step 4: Define a function to handle mouse events:   9: functiononMouse(event, x, y, flags, param)   10:     if event is left mouse button down then   11:         Capture starting coordinates of ROI (x, y)   12:     end if   13:     if event is left mouse button up then   14:         Capture ending coordinates of ROI (x, y)   15:         Draw rectangle on the frame to visualize the ROI   16:         Display the frame with ROI   17:     end if   18: end function   19: Step 5: Create a named window for displaying the video frame.   20: Step 6: Register the mouse callback function to the named window created in step 5.   21: Step 7: Loop over frames:   22: while video is not ended do   23:     Read the next frame from the video capture object.   24:     Display the frame in the named window.   25: end while   26: Step 8: Check for key press events:   27: if key is pressed then   28:     if key is designated key for saving ROI then   29:         Save the defined ROI as an image file.   30:     end if   31: end if   32: Step 9: Release the video capture object and close all OpenCV windows.

6. Results

6.5. F1-Score:

The F1-Score offers a balanced evaluation by calculating the harmonic mean of precision (Equation (4)) and recall (Equation (5)). This metric provides a robust assessment of model performance, less susceptible to outliers compared to individual precision and recall values. It is expressed as:

$$\begin{array}{cc}\hfill F1-Score& =\frac{2\times \mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}\hfill \end{array}$$

(8)

Valuable insights into the strengths and weakness of the YOLOv8 models for the vehicle and license plate detection tasks within our smart parking system are gained through and in-depth analysis of these evaluation metrics. This analysis facilitates informed decision-making regarding the most suitable YOLOv8 variant for real-world application deployment.

In Table 1, a comprehensive comparison of a performance metrics for various YOLOv8 variants for vehicle detection and license plate detection respectively, is presented. The evaluation encompasses precision (Equation (4)), recall (Equation (5)), mean Average Precision (mAP) (Equation (7)) at 50% and 95% IoU (Equation (6)) thresholds, and F1 scores (Equation (8)).

Table 1. Comparison of metrics for Vehicle and License Plate Identification of different YOLO Models

Table 1. Comparison of metrics for Vehicle and License Plate Identification of different YOLO Models

	Vehicle Detection					License Plate Detection
Model	Precision	Recall	mAP50	mAP50-95	F1 Score	Precision	Recall	mAP50	mAP50-95	F1 Score
Nano	0.847	0.840	0.815	0.719	0.843	0.949	0.848	0.916	0.711	0.896
Small	0.863	0.843	0.821	0.747	0.852	0.942	0.853	0.916	0.718	0.895
Medium	0.864	0.849	0.853	0.768	0.856	0.942	0.86	0.919	0.716	0.899
Large	0.874	0.854	0.849	0.787	0.864	0.946	0.859	0.919	0.721	0.900

7. Analysis

7.1. TOPSIS Analysis:

The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) [22,23] is a method for Multiple Criteria Decision Making (MCDM), initially conceived by Ching-Lai Hwang and Yoon in 1981, with subsequent refinements by Yoon in 1987 and Hwang, Lai, and Liu in 1993. TOPSIS operates on the principle that the preferred alternative should be closest to the positive ideal solution (PIS) and farthest from the negative ideal solution (NIS) in geometric distance.

7.1.1. Procedure

The TOPSIS process is carried out as follows:

• Step 1: Create an evaluation matrix consisting of m alternatives and n criteria, with the intersection of each alternative and criteria given as $x_{ij}$. Therefore, we have a matrix ${\left(x_{ij}\right)}_{m\times n}$.
• Step 2: The matrix ${\left(x_{ij}\right)}_{m\times n}$ is then normalized to form the matrix

  $$R={\left(r_{ij}\right)}_{m\times n},$$

  (9)

  using the normalization method

  $$\begin{array}{cc}\hfill r_{ij}& =\frac{x_{ij}}{\sqrt{{\sum }_{k=1}^m{x}_{kj}^2}},\hfill \\ \hfill & i=1,2,\dots ,m,j=1,2,\dots ,n.\hfill \end{array}$$

  (10)
• Step 3: Calculate the weighted normalized decision matrix

  $$t_{ij}=r_{ij}·w_j,i=1,2,\dots ,m,j=1,2,\dots ,n$$

  (11)

  where $w_j=W_j/{\sum }_{k=1}^n{W}_k,j=1,2,\dots ,n$ so that ${\sum }_{i=1}^n{w}_i=1$, and $W_j$ is the original weight given to the indicator $v_j,j=1,2,\dots ,n.$
• Step 4: Determine the worst alternative $\left(A_w\right)$ and the best alternative $\left(A_b\right)$:

  $$\begin{array}{c}\hfill \begin{array}{cc}\hfill A_w& =\left\{\langle max(t_{ij}\mid i=1,2,\dots ,m)\mid j\in J_{-}\rangle ,\right.\hfill \\ \hfill & \left.\langle min(t_{ij}\mid i=1,2,\dots ,m)\mid j\in J_{+}\rangle \right\}\hfill \\ \hfill & \equiv \{t_{wj}\mid j=1,2,\dots ,n\},\hfill \\ \hfill A_b& =\left\{\langle min(t_{ij}\mid i=1,2,\dots ,m)\mid j\in J_{-}\rangle ,\right.\hfill \\ \hfill & \left.\langle max(t_{ij}\mid i=1,2,\dots ,m)\mid j\in J_{+}\rangle \right\}\hfill \\ \hfill & \equiv \{t_{bj}\mid j=1,2,\dots ,n\},\hfill \end{array}\end{array}$$

  where,

  $$J_{+}=\{j=1,2,\dots ,n\mid j\}$$

  associated with the criteria having a positive impact, and

  $$J_{-}=\{j=1,2,\dots ,n\mid j\}$$

  associated with the criteria having a negative impact.
• Step 5: Calculate the $L^2$-distance between the target alternative i and the worst condition $A_w$

  $$d_{iw}=\sqrt{\sum _{j=1}^n{(t_{ij}-t_{wj})}^2},i=1,2,\dots ,m,$$

  (12)

  and the distance between the alternative i and the best condition $A_b$

  $$d_{ib}=\sqrt{\sum _{j=1}^n{(t_{ij}-t_{bj})}^2},i=1,2,\dots ,m,$$

  (13)

  where $d_{iw}$ and $d_{ib}$ are $L^2$-norm distances from the target alternative i to the worst and best conditions, respectively.
• Step 6: Calculate the similarity to the worst condition:

  $$s_{iw}=\frac{d_{iw}}{d_{iw}+d_{ib}},0\le s_{iw}\le 1,i=1,2,\dots ,m.$$

  (14)
  $s_{iw}=1$ if and only if the alternative solution has the best condition; and $s_{iw}=0$ if and only if the alternative solution has the worst condition.
• Step 7: Rank the alternatives according to $s_{iw}(i=1,2,\dots ,m)$.

An open-source Python program [24] has been utilized for the TOPSIS analysis. Our input matrix consists of 7 parameters, namely Precision(Equation (4)), Recall(Equation (5)) and mAP50-95(Equation (7)) for each of the two models (Vehicle and License plate Detector) along with Total Inference Time in milliseconds.

7.1.2. Criteria Weighting

Now the assignment of weights, $W_j$ must be done on the basis of priority given to each metric for our specific application. Here’s how the optimization of the combination of 16 different types of system design must be carried out:

• Precision (Equation (4)): It becomes paramount for both car and license plate detection. False positives, which can include detecting multiple cars, non-car objects, or non-plate objects, are highly detrimental as they hinder accurate identification.
• mAP50-95 (Equation (7)): As IoU remains crucial for both car and license plate detection, mAP at 95% IoU threshold has been considered as the one of the criteria in TOPSIS. Also, accurate bounding box localization is essential for reliable downstream tasks like Optical Character Recognition (OCR) applied to license plates.
• Recall (Equation (5)): Recall can be slightly relaxed for car detection compared to a multi-car scenario. Missing a single car may be less frequent. However, a decent recall rate is still desirable. Recall becomes less critical for license plate detection compared to car detection.

The weights $W_i$ for both the models’ metrics are as follows:

• Precision: 20
• mAP50-95: 15
• Recall: 10

Weight $W_i$, of 1 has been assigned to Total Inference Time, in order to balance the trade-off between speed and accuracy.

7.1.3. Results

In Table 2, according to TOPSIS analysis, top 2 model combinations(highlighted in bold) are:

• TOPSIS Rank 1: YOLOv8s (Small) for Vehicle detection + YOLOv8m (Medium) for License Plate detection
• TOPSIS Rank 2: YOLOv8m (Medium) for Vehicle detection + YOLOv8m (Medium) for License Plate Detection.

Next, light intensity analysis will be performed for the top-ranked models according to TOPSIS analysis.s

Table 2. Performance Comparison of Car Detector and License Plate Detector

Table 2. Performance Comparison of Car Detector and License Plate Detector

S.NO.	VEHICLE DETECTION				LICENSE PLATE DETECTION				TOTAL INFERENCE TIME	FPS	TOPSIS
	Model	Precision	Recall	mAP50-95	Model	Precision	Recall	mAP50-95	PER FRAME (Milliseconds)		Rank
1	NANO	0.847	0.84	0.815	NANO	0.949	0.848	0.719	3.8	263.15	15
2					SMALL	0.942	0.853	0.747	6.3	158.73	7
3					MEDIUM	0.942	0.86	0.768	12.2	81.96	5
4					LARGE	0.946	0.859	0.787	19.2	52.08	8
5	SMALL	0.863	0.843	0.821	NANO	0.949	0.848	0.719	6.3	158.73	11
6					SMALL	0.942	0.853	0.747	8.8	113.63	3
7					MEDIUM	0.942	0.86	0.768	14.7	68.02	1
8					LARGE	0.946	0.859	0.787	21.7	46.08	6
9	MEDIUM	0.864	0.849	0.853	NANO	0.949	0.848	0.719	12.2	81.96	14
10					SMALL	0.942	0.853	0.747	14.7	68.02	4
11					MEDIUM	0.942	0.86	0.768	20.6	48.54	2
12					LARGE	0.946	0.859	0.787	27.6	36.23	9
13	LARGE	0.874	0.854	0.849	NANO	0.949	0.848	0.719	19.2	52.08	16
14					SMALL	0.942	0.853	0.747	21.7	46.08	12
15					MEDIUM	0.942	0.86	0.768	27.6	36.23	10
16					LARGE	0.946	0.859	0.787	34.6	28.9	13

7.2. Light Intensity Analysis:

For simulating the varying lighting conditions, a controllable light source along with an android application [25] has been used to measure the light intensity at the region of interest. A smartphone camera, coupled with Camo Studio [26], has been utilized to ensure high-quality footage and integration with OpenCV library. The camera specifications [27] are as follows:

• Main Camera: 12-megapixel (12MP)
• Aperture: f/1.8
• Lens Configuration: 5P (5-element lens)
• Autofocus Technology: Phase Detection Autofocus (PDAF)
• Video Output: 1080p at 30FPS with Electric image stabilization (EIS)

A model car and the Indian High Security Registration Plate (HSRP) of 1:18 scale have been utilized to ensure size continuity in the simulation environment.

Figure 12. Variation of light intensity.

Figure 12. Variation of light intensity.

According to Indian Standards [28], minimum level of luminance on the roads must be 30 lux, in order to ensure safety of the commuters. Thats’s why the ambient lighting level has been reduced to 20 lux in the analysis shown in Table 3. This proves an estimate of how well the system perform in low lighting conditions.

8. Discussion

9. Conclusion and Future Scope

References

1. McCoy K (2023) Drivers spend an average of 17 hours a year searching for parking spots. 2017.
2. Cookson G, Pishue B (2017) Searching for parking costs americans $73 billion a year.
3. Jocher G, Chaurasia A, Qiu J (2023) Ultralytics yolov8. https://github.
4. Wojke N, Bewley A, Paulus D (2017) Simple online and realtime tracking with a deep association metric. In: 2017 IEEE International Conference on Image Processing (ICIP), IEEE, pp 3645–3649, 10.1109/ICIP.2017.8296962.
5. Wojke N, Bewley A (2018) Deep cosine metric learning for person re-identification. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), IEEE, pp 748–756, 10.1109/WACV.2018.00087.
6. Cao J, Pang J, Weng X, et al (2023) Observation-centric sort: Rethinking sort for robust multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 9686–9696.
7. Sharma N, Baral S, Paing MP, et al (2023) Parking time violation tracking using yolov8 and tracking algorithms. Sensors 23(13):5843.
8. Ultralytics (2022) ultralytics/yolov5: v7.0 - YOLOv5 SOTA Realtime Instance Segmentation. https://github.com/ultralytics/yolov5.com, 10.5281/zenodo.7347926, https://doi.org/10.5281/zenodo.7347926, accessed: 7th May, 2023.
9. Nguyen DL, Vo XT, Priadana A, et al (2023) Car detection for smart parking systems based on improved yolov5. Vietnam Journal of Computer Science 1:15.
10. Hasan Yusuf F, A Mangoud M (2024) Real-time car parking detection with deep learning in different lighting scenarios. International Journal of Computing and Digital Systems 15(1):1–9.
11. Das S (2019) A novel parking management system, for smart cities, to save fuel, time, and money. In: 2019 IEEE 9th annual computing and communication workshop and conference (CCWC), IEEE, pp 0950–0954.
12. Redmon J, Divvala S, Girshick R, et al (2016) You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 779–788.
13. Badla S (2014) Improving the efficiency of tesseract ocr engine.
14. n (2023) cars dataset. https://universe.roboflow.com/n-b8dyr/cars-ydptb , https://universe.roboflow.com/n-b8dyr/cars-ydptb, visited on 2024-03-30.
15. projects M (2024) license plates dataset. https://universe.roboflow.com/merge-projects/license-plates-igyvb , https://universe.roboflow.com/merge-projects/license-plates-igyvb, visited on 2024-03-30.
16. Smith R (2007) An overview of the tesseract ocr engine. In: Ninth international conference on document analysis and recognition (ICDAR 2007), IEEE, pp 629–633.
17. OpenCV (2015) Open source computer vision library.
18. Clark A (2015) Pillow (pil fork) documentation. https://buildmedia.readthedocs.org/media/pdf/pillow/latest/pillow.pdf.
19. Ren S, He K, Girshick R, et al (2016) Object detection networks on convolutional feature maps. IEEE transactions on pattern analysis and machine intelligence 39(7):1476–1481.
20. Sawant M (2022) Pytesseract: Python-tesseract is an optical character recognition (ocr) tool for python. that is, it will recognize and "read" the text embedded in images. https://pypi.org/project/pytesseract/.
21. Kim J, Cho J (2021) Rgdinet: Efficient onboard object detection with faster r-cnn for air-to-ground surveillance. Sensors 21(5):1677.
22. Hwang CL, Yoon K (2012) Multiple attribute decision making: methods and applications a state-of-the-art survey, vol 186. Springer Science & Business Media.
23. Zavadskas EK, Zakarevicius A, Antucheviciene J (2006) Evaluation of ranking accuracy in multi-criteria decisions. Informatica 17(4):601–618.
24. Wikipedia (2024) TOPSIS — Wikipedia, the free encyclopedia. http://en.wikipedia.org/w/index.php?title=TOPSIS&oldid=1185091184, [Online; accessed 02-April-2024].
25. Phuongpn (2023) Illuminance - lux light meter. https://phuongpndev.web.app/, [Accessed 02-04-2024].
26. Ltd. R (2020) Camo studio. https://reincubate.com/camo/, [Accessed 02-04-2024].
27. (2019) Realme 5 Smartphone. Realme Mobile Telecommunications (India) Private Limited.
28. Institution IS (1981) Indian Standard CODE OF PRACTICE FOR LIGHTING OF PUBLIC THOROUGHFARES. Indian Standards Institution, New Delhi, iS 1944-5 (1981).
29. Wang CY, Liao HYM (2024) YOLOv9: Learning what you want to learn using programmable gradient information.