A Smart Car Parking System Based on IoT with Gray Wolf Optimization-Probability Correlated Neural Network Recognition Methods

In the recent days, the Internet of Things (IoT) [1, 2] technology has gained a significant attention in different types of application systems, due to its benefits of high system efficiency, cost-effective nature, and better opportunities. Also, the utilization rate of mobiles, wireless sensors, and embedded system are rapidly increased with the use of the IoT [3-6]. The smart car parking system is one of the interesting and demanding research field in smart city systems [7, 8]. The main purpose of developing the smart car parking system is to avoid the traffic congestions and time delay of processing and waiting. Typically, the smart car parking system [9] has the following benefits:

1.       It exactly predicts and sense the occupancy space of vehicles for avoiding congestion.

2.       It optimally utilizes the parking space of all vehicles.

3.       Due to the incorporation of IoT technology, it efficiently avoids traffic in the city.

4.       By using the soft computing methodologies, intelligent decisions are taken at the complex situations.

5.       Moreover, the smart car parking system could efficiently create the urban environment with reduced Co₂ emission and other pollutants.

Due to this facts, various car parking architectures have been developed in the conventional works [10-12], which utilizes some machine learning or deep learning models for improving the performance of recognition. In the proposed work, the image processing techniques are employed for accurately recognizing the car image with desired parking location [13, 14]. The major contributions of this research work are as follows:

·         To develop an advanced car parking system for reducing the traffic congestion with the help of IoT technology.

·         To accurately recognize the parking area, the image preprocessing system is incorporated, which helps to predict whether the parking area is available or not.

·         To efficiently preprocess the given image, an Anisotropic Diffusion Gaussian Filtering (ADGF) technique is employed that eliminates the noisy content with increased quality.

·         To extract the set of features from the normalized image, the Grey Level Co-occurrence Matrix (GLCM) technique is utilized.

·         To optimally select the best features based on the optimal fitness function, the Grey Wolf Optimization (GWO) technique is employed.

·         To improve the recognition performance, the Probability Correlated Neural Network (PCNN) technique is deployed.

The remaining portions of this paper are segregated into the following sections: the existing works related to the car parking system in IoT-cloud environment are surveyed with its advantages and disadvantages in Section II. Then, the description of the overall proposed methodology with its appropriate algorithmic illustrations are presented in Section III. Moreover, the overall performance results and comparative analysis of the existing and proposed security models are illustrated in Section IV. Finally, the entire paper is summarized with its advantages and future scope in Section V.

This section presents the detailed description of the proposed methodology with its overall architecture and algorithmic illustrations. The main contribution of this paper is to develop an efficient car parking system incorporated with the IoT technology for optimally allocating parking area. Here, the car image identification is performed for accurately recognizing the car located in the specified area by using the optimization based machine learning model. Also, an IoT based monitoring and controlling unit is utilized in this framework for reducing the waiting time, delay of process, and response time. In the car parking scenario, a group of users are required to spend their time for parking the cars in the desired locations. The proposed model objects to reduce the queuing time with the use of car parking management scheme. The overall architecture representation of the proposed model is shown in Figure 1, where IoT cloud server is deployed for notifying the information of cars to the users. When the user parks the car, the request has been automatically generated and passed to IoT server unit. Then, the controlling unit instructs the camera to capture the corresponding car image for monitoring the desired parking location of user. This type of monitoring system helps the user to easily find their vehicle in the parking place with reduced queuing time and delay time. Furthermore, the optimization based machine learning methodology is used for recognizing that whether the destination parking area is available or not. The overall flow of the car image processing system is shown in Figure 2, which includes the following stages:

• Input image obtainment
• Image preprocessing using Anisotropic Diffusion Gaussian Filtering (ADGF) Feature extraction using GLCM
• Grey Wolf Optimization
• Probability Correlated Neural Network (PCNN) based recognition

1.png

Figure 1. Architecture model of the proposed system

2.png

Figure 2. Flow of the proposed system

3.1 Image preprocessing

Generally, image normalization is one of the essential process for the recognition system, because the accuracy of classification is highly depends on the quality of image. Hence, the input image must be preprocessed for eliminating the noise contents, reducing the blurring effects, and increasing the contrast or quality of image. Here, the Anisotropic Diffusion Gaussian Filtering (ADGF) technique is utilized to preprocess the input car image by increasing its contrast and quality of image. Typically, the anisotropic diffusion filters are mainly used for reducing the noise with better edge preservation. Also, it helps to strengthen the edges by solving the scale space problem based on diffusion. Thus, the proposed work intends to use an enhanced ADGF technique for enhancing the overall quality of input image. Here, the diffusion coefficient $\mathrm{d}_{\mathrm{c}}$ is selected for sharping the boundaries with inter-region smoothing. For the input car image $\mathrm{I}_{\mathrm{c}}$, the anisotropic diffusion is computed as follows:

$\mathrm{I}_{\mathrm{cx}}=\operatorname{div}\left(\mathrm{d}_{\mathrm{c}}(\mathrm{a}, \mathrm{b}, \mathrm{x}) \nabla \mathrm{I}_{\mathrm{c}}\right)=\mathrm{d}_{\mathrm{c}}(\mathrm{a}, \mathrm{b}, \mathrm{x}) \Delta \mathrm{I}_{\mathrm{c}}+\nabla \mathrm{d}_{\mathrm{c}} \cdot \nabla \mathrm{I}_{\mathrm{cx}}$         (1)

where, $d_c(a, b, x)$ indicates the conduction coefficient, div denotes the divergence operator, $\Delta$ is the gradient operator and $\nabla$ represents the Laplacian operator. In which, the conduction coefficient is computed for properly computing the edges of image as represented in below:

$\mathrm{d}_{\mathrm{c}}(\mathrm{a}, \mathrm{b}, \mathrm{x})=\mathrm{s}\left(\left\|\nabla \mathrm{I}_{\mathrm{c}}(\mathrm{a}, \mathrm{b}, \mathrm{x})\right\|\right)$        (2)

where, $I_{c x}=d_c \Delta d_c$ and its diffusion coefficient is $\left(s\left(\nabla I_c\right)\right)$. After that, the Gaussian kernel function has been utilized to estimate the diffusion coefficients as illustrated in below:

$s\left(\left\|\nabla\left(G \times I_c\right)\right\|\right)$      (3)

Based on this value, the obtained anisotropic diffusion Gaussian function is computed as follows:

$I_c=\operatorname{div}\left(\beta . d_c\left(\left\|\nabla I_c\right\|\right) \cdot \nabla I_c\right)$         (4)

$\beta=1-\nabla G_\sigma \times d_c\left(I_c\right)$          (5)

where, $\beta$ indicates the adaptive term, and $G_\sigma$ denotes the Gaussian kernel. By using function, the image is normalized with the diffusion coefficients and, the preprocessed image can be utilized for further processes.

3.2 Gray Level Co-occurrence Matrix (GLCM) feature extraction

After preprocessing, the Gray Level Co-occurrence Matrix (GLCM) based feature extraction methodology is used for extracting the set of features from the normalized image. It is one of the extensively used texture based extraction model, where the relationship between the pixels are estimated according to the second-order statistical information. In this framework, four different types of GLCM features such as contrast, correlation, energy and homogeneity have been computed as shown in below:

$\operatorname{Con}_T=\sum_{x, y}|x-y|^2 p(x, y)$         (6)

$\operatorname{Cor}_r=\sum_{x, y} \frac{(x-\mu x)(y-\mu y) p(x, y)}{\sigma x \sigma y}$        (7)

$E n_g=\sum_{x, y} p(x, y)^2$         (8)

$\mathrm{Hom}_g=\sum_{x, y} \frac{p(x, y)}{1+|x-y|}$        (9)

where, x indicates the row of image, y indicates the column of image, and p is the total number of pixels in the image. In which, the contrast is defined as the color difference value that is mainly extracted for determining the objects of image. Then, the correlation is defined as the kind of dissimilarity measure that is mainly computed for analyzing the contrast of texture value. Similar to that, the energy is computed according to the global uniformity of the given image, and the homogeneity is estimated based on the proximity of element distribution. These features are more useful for improving the recognition rate of classifier.

3.3 Grey Wolf Optimization (GWO) based feature selection

After extracting the Gray Level Co-occurrence Matrix (GLCM) features, the Grey Wolf Optimization (GWO) based feature selection mechanism is utilized for selecting the best suitable features based on the optimal solution. It is a kind of meta-heuristic optimization technique and mainly used for solving complex optimization problems. The major benefits of using this technique are as follows: Increased convergence speed, best optimal solution with reduced number of iterations, and reduced complexity. Hence, the proposed work intends to utilize GWO technique for optimally selecting the features of preprocessed car image, which helps to train the classifier for improving the recognition rate. In this technique, the grey wolves are segregated into four different groups such as $\alpha, \beta, \delta$, and $\omega$, in which alpha, beta, and delta are considered as the most essential gray wolves. In addition to that, it effectively identifies the feature space for selecting the best subset of feature according to the fitness value. Here, the encircling behavior of grey wolves are computed as follows:

$\overrightarrow{\mathrm{GV}}(\mathrm{a}+1)=\overrightarrow{\mathrm{GV}}(\mathrm{a})+\overrightarrow{\mathrm{X}} \cdot \overrightarrow{\mathrm{Y}}$        (10)

where, $\overrightarrow{\mathrm{GV}}$ indicates the prey’s position vector, a is the number of iterations, $\overrightarrow{\mathrm{GV}_{\mathrm{p}}}$ defines the vector of grey wolf, and $\vec{X}, \vec{Y}$, and $\overrightarrow{\mathrm{Z}}$ are the coefficient vectors computed as follows:

$\overrightarrow{\mathrm{Z}}=\left|\overrightarrow{\mathrm{Y}} \cdot \overrightarrow{\mathrm{GV}_{\mathrm{P}}}(\mathrm{a})-\mathrm{GV}(\mathrm{a})\right|$       (11)

Similar to that, the coefficients $\overrightarrow{\mathrm{X}}, \overrightarrow{\mathrm{Y}}$ are estimated as shown in below:

$\overrightarrow{\mathrm{X}}=2 \overrightarrow{\mathrm{v}} \cdot \overrightarrow{\mathrm{t}_1}-\overrightarrow{\mathrm{v}}$        (12)

$\vec{Y}=2 \cdot \overrightarrow{t_2}$        (13)

where, $\overrightarrow{\mathrm{v}}$ indicates the vector value, $\overrightarrow{t_1}$ and $\overrightarrow{t_2}$ are the random values. Consequently, the best fitness value of grey wolves $\alpha, \beta, \delta$, and $\omega$ are computed as follows:

Fitness $=\alpha \mathrm{C}+\beta \frac{\mathrm{No}_{\mathrm{F}}-\mathrm{L}_{\mathrm{F}}}{\mathrm{L}_{\mathrm{F}}}$        (14)

where, $\mathrm{No}_{\mathrm{F}}$ indicates the number of features, and $\mathrm{L}_{\mathrm{F}}$ is the length of feature. The algorithmic steps involved in GWO technique are illustrated in below:

Input:     Extracted set of features, and N_n number of searching agents;

Output:  Optimal solution Opt_B;

Step 1:   Initialize the all populations of searching agents randomly as shown in below:

                S_i (i=1,2…n); //n-dimension

Step 2:   Initialize the optimization variables as k, K and L;

Step 3:   Compute the fitness of each search agent as illustrated below:

              S_α = Best search agent;

              S_β = Second best;

              S_δ = Third best;

Step 4:      while (until reaching the stopping criterion) do

Step 5:      for i = 1 to N do

         The position of S_i is updated;

         end for;

Step 6:      Update the positions of k, K and L;

Step 7:      Estimate the fitness value;

Step 8:      Update the parameters as, S_α, S_β, and S_δ;

Step 9:      end while;

Step 10:    Return Opt_B=S_α;

After optimization, the best suited features are considered as the input for classification, which is one of the essential stage in the recognition systems. When compared to the other classification approaches like KNN, DT, NB, C4.5, SVM, RVM, DNN, DBN, and LSTM. Here, the main purpose of using the classification technique is to accurately predict the desired car location with reduced time consumption. It is a kind of machine learning technique, which accepts the features as the input and predicted/recognized label as the output. In this algorithm, the input layers comprises the samples of $a_i=a_1, a_2 \ldots a_N$, where $a_i$ indicates the input vector, and it generates the output function as follows: $\mathrm{b}_{\mathrm{i}}=\mathrm{b}_1, \mathrm{~b}_2 \ldots \mathrm{b}_{\mathrm{N}}$. The algorithmic steps involved in this work are as follows:

Step 1:   The center value is computed by estimating the distance value as shown in below:

         $\mathrm{Di}_{\mathrm{s}}=\sum_{\mathrm{y}=1}^{\mathrm{n}} \exp \left(-\frac{\left\|\mathrm{a}_{\mathrm{i}}-\mathrm{p}_{\mathrm{i}}\right\|}{\left(\mathrm{Nei}_{\mathrm{r} / 2}\right)^2}\right)^2$

Step 2:      $D i_s-D i_{t p} \exp \left(\frac{\| S_s-s_p}{\left(N^{-i} i_{r / 2}\right)^2}\right)^2$

Step 3:   For all input samples, the distance value is computed as follows:

                $\mathrm{F}_{\mathrm{xy}}=\left\|\mathrm{a}_{\mathrm{j}}-\mathrm{p}_{\mathrm{i}}\right\|$

Step 4:   Correspondingly, the minimum distance value is estimated as follows:

                $\mathrm{p}_{\mathrm{i}}^{\prime}=\frac{1}{\mathrm{~h}_{\mathrm{i}}} \sum_{\mathrm{a} \in \mathrm{c}_{\mathrm{i}}} \mathrm{a}(1 \leq \mathrm{b} \leq \mathrm{m})$

Step 5:   Then, the cluster center is estimated and position updation are performed, and the output associates with the hidden layer could produce the recognized results.

This paper developed a new smart car parking system by using an IoT technology for avoiding the traffic congestion in the smart cities. The main purpose of this work was to develop an IoT based monitoring and controlling unit for reducing the waiting time, delay of process, and response time. Then, the controlling unit instructs the camera to capture the corresponding car image for monitoring the desired parking location of user. This type of monitoring system helps the user to easily find their vehicle in the parking place with reduced queuing time and delay time. Here, the Anisotropic Diffusion Gaussian Filtering (ADGF) technique is utilized to preprocess the input car image by increasing its contrast and quality of image. The proposed work intends to utilize the Grey Wolf Optimization (GWO) technique for optimally selecting the features of preprocessed car image, which helps to train the classifier for improving the recognition rate. The main purpose of using the PCNN classification technique is to accurately predict the desired car location with reduced time consumption. During evaluation, the performance of the proposed technique is validated and compared by using various measures. From the obtained results, it is concluded that the proposed GWO based PCNN technique outperforms the other techniques by accurately recognizing the car image with reduced error rate.