A novel traffic optimization method using GRU based deep neural network for the IoV system

Deep learning

As one of the branches of machine learning, the concept of deep learning can be traced back to 2006, when it was introduced by the scholar G. E. Hinton. Unlike traditional machine learning methods such as SVM (Aurangzeb, Ayub & Alhussein, 2021), Boosting (Konstantinov & Utkin, 2021), and SAE (Ji, Zhang & Wu, 2020), deep learning utilizes a deep network architecture containing multiple layers to perform some processing on the input data in turn, while initializing the neural network parameters using unsupervised pre-training methods. By training the model iteratively, the layers of the network become deeper and deeper, and the original data are trained to different degrees, gradually making the training results meet the feature representation required for the target task. Finally, the extracted data features are classified and regressed to produce the results (Naghizadeh, Metaxas & Liu, 2021). The deep learning can directly process the original data at multiple levels, effectively avoiding the process of preprocessing complex data with traditional machine learning models. Thus, manual intervention is reduced, and its prediction accuracy and operation efficiency are greatly improved. The core component of deep learning models is a large number of neuron structures that hierarchically perform feature extraction on original data through multiple transformation stages and interpret these data  (Ke et al., 2021). Figure 1 displays the data processing flow.

Neurons

The most fundamental components of a deep neural network are neurons  (Bodyanskiy & Antonenko, 2021). Numerous neurons can be combined to form a complicated deep neural network. They are connected according to certain rules and then distributed in each layer of the network. In general, neurons may consist of either linear products or activation functions using nonlinear parts. Figure 2 depicts its structure.

Usually, each neuron typically does not exist in isolation in deep neural network models. On the other hand, they simultaneously connect to a number of other neurons in different network layers. In this way, the output signals sent from other neurons can be used as input signals for that neuron. Suppose the input signal of each neuron is defined as S =(s ₁,s ₂, …., s_m,), and its corresponding weight matrix is V =(v ₁ ,v ₂ , …., v_m). Then the output signal of this neuron can be expressed by Eq. (1). (1)${\displaystyle z=f\left(\sum _{i=1}^m{v}_i{s}_i+a\right)}$ where f represents the activation function and a means the bias vector. v and a are the key variables of the neuron and will significantly affect the output of the neuron. During the model training process, these two variables are continuously adjusted to achieve model optimization. The activation function f also plays a pivotal role in the neuron’s processing of the input data and is one of the key aspects of the model training. If the neuron consists of only linear functions, then the output results obtained through model training are only linear transformations of the input data, which greatly limits the learning ability of the model. Therefore, the nonlinear part of the activation function f is introduced into the neurons, so as to enhance the nonlinear mapping learning ability of deep neural networks. Since the characteristics of different activation functions are also different, their impact on the model training results is not exactly the same. Currently, the commonly used activation functions are sigmoid function and tanh function. The former has the advantage of being easy to derive and the disadvantage of having problems such as gradient disappearance. The latter has the advantage of faster convergence.

Deep feedforward network

The deep feedforward network is one of the most well-known artificial neural networks, which has a typical deep neural network structure  (Kozlenko, Zamikhovska & Zamikhovskyi, 2021). It is usually a one-way multilayer network structure with no signal feedback consisting of input, hidden and output layers, so its whole network can be expressed as a directed acyclic graph. Moreover, the network contains a large number of multilayer neurons. The data is fed into the network’s previous layer’s neurons for processing, and after the processing is finished, the output is fed into the network’s next layer’s neurons. Figure 3 depicts the structure of the deep feedforward network.

In Fig. 3, the deep feed-forward neural network uses forward propagation to calculate the output of the model. The input layer is located in the first layer of the entire network model and takes on the role of receiving external data signals. The hidden layer usually contains numerous multilayer neurons. Moreover, the fitting degree of the model is often related to the number of layers in the hidden layer. The more layers, the better the fitting degree of the model, which is more conducive to solving complex problems. Meanwhile, it is important to note that if the network has too many hidden layers, it will lead to an increase of neuron parameter variables in the network, which will increase the model training burden and cause problems such as gradient stability and network degradation. The output layer is at the end of the network structure, whose main task is to output the training results of the model.

Error back propagation

The artificial neural network’s scale is substantially increased by the addition of multi-level hidden layers, which also improve the model’s capacity for feature representation. To ensure that the final output results match the feature representation required by the target task, it is important to note that the network’s parameters must be adjusted repeatedly throughout the model training process. However, as the network model’s layers get deeper and deeper, this could also strain the model and reduce its effectiveness. One of the frequently utilized algorithms for training neural networks at the present is the error back propagation algorithm. It can effectively reduce the number of solution parameter gradient expressions, so as to quickly adjust the network model parameters. Figure 4 depicts its training process.

As illustrated in Fig. 4, the output data from the forward propagation in the neural network serves as the input data in the error back propagation algorithm. It uses the loss function to calculate the error of the model prediction, and then the error signal is backward propagated through the network layer by layer. It also utilizes the chain rule to calculate the gradient of the weights of each neuron and updates the weights layer by layer using the gradient descent algorithm.

Deep learning-based traffic flow prediction model

This study optimizes the GRU neural network model starting from the direction of road network traffic optimization in the IoV environment. In this manner, the generalizability and prediction accuracy of current traffic flow prediction techniques will be increased. The prediction model mainly includes the steps of traffic data collection, prediction model training, result analysis and traffic alert. Figure 5 displays the detailed workflow of the model.

Figure 5 shows the detailed workflow of the model. (1) Traffic data collection. Traffic data such as weather, vehicle type, number of vehicles, and so on are collected in real time using road measurement sensors, cameras, and other related devices in the IoV environment. (2) Model training. The collected traffic data are pre-processed and then directly used as input data for the prediction model for model training. (3) Results analysis. The prediction results of the model are transmitted to the traffic command center using advanced network communication technology. (4) Traffic alert. The traffic command center further analyzes and calculates the prediction results in depth and feeds the results to the relevant vehicle units, thus providing reasonable suggestions for people’s travel arrangements. The purpose of road traffic optimization is finally realized.

GRU module design

The article chooses to use the GRU (gated recurrent unit) model to build a deep neural network for short-time prediction of traffic flow in the IoV environment. The GRU neural network is a model improved on the basis of LSTM by Cho et al. (Behera, Misra & Sillitti, 2021). Its performance is close to that of LSTM, but the difference is that it combines the forgetting gate and the input gate in LSTM as a reset gate, so the internal structure of the GRU model is simpler.

A simple GRU model typically only has two gate structures: the update gate and the reset gate. The GRU model has one fewer gate structure than the LSTM model, which results in a decrease in the total number of parameters for the overall model. Figure 6 depicts the internal structure of the GRU model.

In Fig. 6, the two gate structures of the GRU model are update gate n_t and reset gate s_t . The n_t is mainly used to determine the degree of receiving information from the previous moment, and the s_t is mainly utilized to determine the degree of ignoring information from the previous moment. n_t has a larger value indicating a greater degree of receiving information from the previous moment cell, and s_t has a larger value indicating that the state information of the previous moment cell has less influence on the state of the current cell. The expressions of both are shown in Eq. (2) and Eq. (3), respectively.

(2)${\displaystyle n_t=\sigma \left(v_n\cdot \left[o_{t-1},x_t\right]\right)}$ (3)${\displaystyle s_t=\sigma \left(v_s\cdot \left[o_{t-1},x_t\right]\right)}$ where v_n, v_s denote the weight matrix of the two gates respectively, x_t denotes the current cell input, o_t−1 denotes the state information of the cell at the previous moment. o_t denotes the output value of the cell at the current moment, which is defined in Eq. (4) and Eq. (5).

(4)${\displaystyle o_t=\left(1-n_t\right)\cdot o_{t-1}+n_t\cdot o_t^{{}^{\prime }}}$ (5)${\displaystyle o_t^{{}^{\prime }}=tan\mathrm{h}\left(v_{o^{{}^{\prime }}}\cdot \left[s_t\cdot o_{t-1},x_t\right]\right)}$

Since there is one less output gate in the GRU model to control its output, the state information of its cells at all moments is output to the cells at the next moment. This makes GRU eliminate the operation process of multiplying several matrices, which makes the GRU model more advantageous in solving the short-time traffic flow prediction problem. This is exactly the reason why the GRU model is chosen in this article.

Fine-grained traffic flow statistical calculation method

Since the traditional traffic flow statistical calculation method is prone to the problem of missing or wrong detection, a fine-grained traffic flow statistical calculation method is proposed for the constructed deep learning model to better calculate the traffic flow data in the IoV scenario. In this study, the road traffic data collected by roadside sensors and cameras are passed to the GRU prediction model for model training, and the vehicle location and category information of each vehicle is obtained through the model training, and then the extracted information of each vehicle is fed into the fine-grained traffic flow statistical algorithm to realize multi-lane and multi-type traffic flow information statistics. Figure 7 displays the whole flow chart.

It is important to note that the work area must be set for this method. The only vehicles that will be counted are those that appear in the work area,any other vehicles are simply ignored. Figure 8 displays the algorithm’s working schematic.

It is assumed that the bounding box information of the vehicle is determined by the coordinate values of the upper left and lower right corners, which are defined as (a^l, b^l) and (a^r, b^r), respectively. (A, B) represent the coordinates of the upper right corner of the detection area. The upper left coordinate of the virtual coil is(a^u, b^u) and the lower right coordinate is (a^v, b^v). t represents the logical time when the virtual coil is created. The expression of the virtual coil is shown in Eq. (6). (6)${\displaystyle R=\left(a_1^u,b_1^u,a_1^v,b_1^v,t_1\right),\dots ,\left(a_n^u,b_n^u,a_n^v,b_n^v,t_n\right)}$

If the bounding box information of the vehicle can satisfy both the conditions of Eq. (7) and Eq. (8), a virtual coil with that vehicle is established in the detection area.

(7)${\displaystyle B-b^r<D}$ (8)${\displaystyle B-b^r>0}$ where D denotes the distance threshold between the bounding box of vehicle and the detection area. When $\left|B-b^r\right|<D$ is used, each frame needs to determine whether there is a virtual coil corresponding to the bounding box of the current vehicle, whose expression is shown in Eq. (9). (9)${\displaystyle \frac{max\left(a^v,a^r\right)-min\left(a^u,a^l\right)}{a^v-a^u}>L}$ where L denotes the threshold value and L ∈ [0, 1] . The above formula indicates that after the current vehicle width is projected onto the virtual coil width, only when the ratio of the projected value to the virtual coil width is greater than the threshold L, it is determined that the current vehicle has a virtual coil corresponding to it. When the bounding box of the vehicle enters into the virtual coil, it is necessary to execute the vehicle counting process at this time, and when the condition of Eq. (10) is satisfied, the vehicle is counted and the current virtual coil is deleted. (10)${\displaystyle \frac{\text{area}\left(F\cap G\right)}{\text{area}\left(G\right)}>H}$ where F represents the rectangular bounding box of the vehicle to be judged, G represents the virtual coil frame to be judged, and H is the vehicle count threshold. At the end, the following situation should also be considered to ensure the robustness of this algorithm, that is, if a certain vehicle creates a certain virtual coil, but for some special reasons, this virtual coil is not deleted after the vehicle passes. In this case, the virtual coil will always exist and may have a negative impact on the later statistics. Therefore, such coils need to be cleared and the virtual coils are deleted when the conditions such as Eq. (11) are met. (11)${\displaystyle t_{\text{now}}-t_{\text{create}}>T_{\max }}$ where T_max is the longest logical time that the virtual coil can be tolerated to exist.

Simulation experiments

The research also performs simulation experiments to confirm the application performance of the traffic flow prediction model designed in this work in the IoV environment. To test its effectiveness, four different scenarios of road videos in varied weather conditions—sunny, cloudy, rainy, and evening—will be constructed using the traffic videos that the camera has captured.

Experimental data

The research in this article focuses on predicting the traffic flow trends of a regional traffic road in the near future based on the traffic data collected in the IoV environment. Therefore, the traffic dataset in this experiment is mainly derived from a hybrid dataset composed of manually labeled dataset and UA-DETRAC public dataset.

The manually labeled dataset contains a total of 1,618 traffic images with a size of 540 × 960, and the shooting scenes involve four situations: sunny, cloudy, rainy and evening. Some of these images were collected by using a camera on a flyover in a certain city, and the other part was obtained by web crawlers. The dataset includes the front and back sides of vehicles. The vehicle types include small cars, buses, cabs, trucks, SUV, and pickup trucks. As a commonly used public vehicle detection dataset, the images within the UA-DETRAC (Perreault et al., 2021) dataset are all collected on the flyover. The dataset collected a total of 82,118 images of 8,250 vehicles in cloudy, evening, sunny, and rainy conditions. The size of each image is 540 × 960 and it involves thirteen types of vehicles, including cabs, cars, SUV, trucks, commercial vehicles and so on. However, since the number of images for each type of vehicles is not equally distributed in the UA-DETRAC dataset, the number of police cars and flatbed trucks is relatively small. In view of this, the article selects the images of nine types of vehicles, including SUV, Sedan, Taxi, Van, MiniVan, Truck-Box, Bus, Truck-Util, and Pickup, from the UA-DETRAC dataset, totaling 11,618. Then the hybrid dataset was formed with 1,618 manually labeled images to ensure the adequacy of computational resources. Then the hybrid dataset was divided into four conditions: sunny day, cloudy day, rainy day, and evening for performance testing. Among them, 10,588 images were selected from the hybrid dataset as the training dataset, and the remaining images were used as the test dataset.

Evaluation indicators

In order to evaluate the performance of the traffic prediction model, this article uses the accuracy rate as the evaluation index. It is defined in Eq. (12). (12)${\displaystyle \text{Accuracy}=\left\{\begin{array}{c}1-\frac{N_{\text{error}}}{N_{\text{true}}},\text{if}{N}_{\text{error}}<N_{\text{true}}\hfill \\ 0,\text{otherwise}\hfill \end{array}\right.}$

It is worth noting that in Eq. (12), if the number of erroneous vehicles exceeds the actual number of vehicles, it exceeds the upper error limit designed in this article. In this case, the prediction accuracy of the model is judged to be 0. Where N_true indicates the number of real traffic flow vehicles and N_error means the number of erroneous vehicles, which is defined in Eq. (13). (13)${\displaystyle N_{\text{error}}=\left|N_{\text{true}}-N_{\text{count}}\right|}$ where N_count represents the number of vehicles counted by the model.