CN111597902A - Motor vehicle illegal parking monitoring method - Google Patents

Abstract

和

3)构建一阶段目标检测算法模型的损失函数LOSS。4)采用梯度下降法对一阶段目标检测算法模型的权值进行更新，直到模型收敛为止；将训练好的模型在实际系统中完成对机动车的检测，并根据现场机动车存在的时间与位置关系，实现对街道内机动车违停的管理。本发明的优点在于能够提高违停车辆监测模型的参数自适应性，大幅提高了违停车辆监测的准确率。The motor vehicle parking violation monitoring method includes the following steps: 1) Collecting a large number of images of high-altitude cameras in the street and other motor vehicle data sets, calibrating the data sets according to on-site management requirements, and determining the one-stage target detection algorithm model to be used. 2) Build a parameter-adaptive loss function

and

3) Construct the loss function LOSS of the one-stage target detection algorithm model. 4) Use the gradient descent method to update the weights of the one-stage target detection algorithm model until the model converges; the trained model is used to detect the motor vehicle in the actual system, and according to the time and location of the motor vehicle on site relationship, and realize the management of illegal parking of motor vehicles in the street. The invention has the advantages that the parameter self-adaptability of the illegally parked vehicle monitoring model can be improved, and the accuracy of illegally parked vehicle monitoring can be greatly improved.

Description

Motor vehicle parking violation monitoring method

technical field

The invention belongs to the technical field of image recognition and computer vision, and relates to a method for monitoring illegal parking of motor vehicles.

Background technique

At present, for the detection of illegal parking of street vehicles, traditional detection methods mainly include: micro-radar detection, infrared detection, geomagnetic induction coil detection and radio frequency identification technology. This type of method requires special sensing equipment to be installed at each location on the street, which has high engineering cost, difficult maintenance in the later stage, and high cost of manpower and material resources. Using the existing security cameras in the street to identify the illegal parking of motor vehicles in the area of the street, no need to change the street surface, and the equipment is easy to maintain and repair, so this video-based motor vehicle parking violation detection system has a good performance. Promote value.

Using the video stream of the security camera to determine whether a motor vehicle is illegally parked in the street area requires high accuracy of the recognition algorithm and real-time information of motor vehicle illegal parking in the application scenario. Therefore, it is more reasonable to use the target detection algorithm based on deep learning. The target detection algorithm based on deep learning is divided into two-stage model and one-stage model. Although the two-stage target detection model has better detection accuracy, its forward reasoning speed is slow and cannot meet the real-time requirements of business scenarios. In the traditional one-stage target detection algorithm model, the real-time performance of the algorithm is good, but the detection accuracy of the two-stage target detection algorithm model cannot be achieved. There are a large number of street background objects in the image detection target. Although the loss value of the street background objects is small, the number is far more than the motor vehicle target. At present, the traditional target detection method is difficult to obtain high recognition in this complex scene. Therefore, a highly adaptive object detection method is urgently needed.

SUMMARY OF THE INVENTION

The present invention aims to overcome the above shortcomings of the prior art, and provides a vehicle parking violation monitoring method with high self-adaptability and high recognition accuracy.

The invention improves the loss function in the one-stage target detection algorithm model. As the objective function of the gradient descent process in the convolutional neural network, the loss function directly affects the training results of the convolutional neural network. The result of convolutional neural network training is directly related to the recognition accuracy of target detection, so the design of loss function is particularly important. In the training process of the one-stage target detection algorithm model, the network contains a large number of street background objects when detecting targets in the image. Although the loss value of the street background objects is small, the number is far more than the motor vehicle target. Therefore, when calculating the loss value, The street background loss value with a small probability value overwhelms the target loss value of the motor vehicle, causing the model accuracy to drop a lot. Therefore, a focal loss function is embedded in the detection model to improve the training accuracy. In the focal loss function, there are hyperparameters that need to be set based on empirical values, and it is impossible to automatically adjust the size of their own hyperparameters based on the predicted class probability value.

Aiming at the problem that the focus loss function needs to manually adjust the hyperparameters in the training process, and the parameters in the training process are not adaptive, the invention proposes a deep learning loss function based on semi-supervised learning. The loss function uses a weighting method to The hyperparameters are improved so that the network can adaptively adjust the network hyperparameters during the gradient descent process, thereby improving the learning efficiency of the network.

The vehicle parking violation monitoring method includes the following steps:

Step 1: Build a motor vehicle sample data set M, a training data set T, a validation data set V, label the number of motor vehicle sample categories C, the training data batch size batch, the training batch number batches, the learning rate l_rate, the training data set T and Scale factor ζ between validation datasets V.

表示图像的高和宽，r表示图像的通道数。Where: V∪T=M, C∈N ⁺ , ζ∈(0,1), batches∈N ⁺ , l_rate∈N ⁺ , batch∈N ⁺ ,

Represents the height and width of the image, and r represents the number of channels in the image.

表示第l层网络中第k个特征图

对应的网格数量，锚点集合M，具体定义如下：Step 2: Determine the one-stage target detection model to be trained, set the depth of the convolutional neural network as L, the network convolution layer convolution kernel set G, the network output layer using the full connection method, the convolution kernel set A, and the network feature map. set U,

Represents the kth feature map in the lth layer network

The corresponding number of grids, the anchor point set M, is specifically defined as follows:

分别表示第l层网络对应的卷积核、特征图和锚点的高、宽、维度。

表示第l层网络卷积核的填充大小，

表示第l层网络卷积步长，f表示卷积神经元的激励函数，Θ表示选取的输入特征，Λ∈N⁺表示第l层网络的锚点总数，Ξ∈N⁺表示输出层节点总数，Φ∈N⁺表示第l层网络特征图总数，Δ∈N⁺表示第l层卷积核的总数。in:

Respectively represent the height, width and dimension of the convolution kernel, feature map and anchor point corresponding to the lth layer network.

Indicates the padding size of the convolution kernel of the lth layer network,

Indicates the convolution step size of the network in the lth layer, f represents the excitation function of the convolution neurons, Θ represents the selected input features, Λ∈N ⁺ represents the total number of anchor points of the network in the first layer, Ξ∈N ⁺ represents the total number of nodes in the output layer , Φ∈N ⁺ denotes the total number of network feature maps in the lth layer, and Δ∈N ⁺ denotes the total number of convolution kernels in the lth layer.

Step 3: Design a focal loss function with adaptive parameters, including:

in:

表示第l层网络上第i个网格中第j个锚点在图像t_k的机动车样本与街道背景样本置信度的损失函数；同理，

表示机动车样本预测框的损失函数，

表示机动车类别的损失函数，λ∈Q为损失函数

参数。

和

分别表示机动车样本目标和街道背景目标的损失函数，具体如下所示：

represents the loss function of the confidence of the vehicle sample and the street background sample of the j-th anchor point in the i-th grid on the l-th network in the image t _k ; similarly,

represents the loss function of the vehicle sample prediction box,

represents the loss function of the vehicle category, λ∈Q is the loss function

parameter.

and

Represent the loss functions of the vehicle sample target and the street background target, respectively, as follows:

表示第l层网络上第i个网格中第j个锚点预测的前景机动车样本概率值，同理，

表示相对应的街道背景概率值。

分别表示第l层网络上第i个网格中第j个锚点的预测框中心点横坐标和纵坐标，同理

分别表示机动车样本标定框的中心点横坐标与纵坐标；

分别表示第l层网络上第i个网格中第j个锚点的预测框中心点到该框边界的最短欧式距离，同理

分别表示机动车样本标定框的中心点到该框边界的最短欧式距离；

表示第l层网络上第i个网格中第j个锚点预测的机动车样本类别预测值。同理，

表示机动车样本类别的标定状态，

表示机动车样本进行预测，

表示是否对街道背景样本进行预测，具体计算如下：

Represents the probability value of the foreground vehicle sample predicted by the jth anchor point in the ith grid on the lth layer network. Similarly,

Represents the corresponding street background probability value.

respectively represent the abscissa and ordinate of the center point of the prediction frame of the jth anchor point in the ith grid on the lth layer network. Similarly,

respectively represent the abscissa and ordinate of the center point of the vehicle sample calibration frame;

Respectively represent the shortest Euclidean distance from the center point of the prediction box of the j-th anchor point in the i-th grid on the l-th network to the boundary of the box. Similarly,

Respectively represent the shortest Euclidean distance from the center point of the vehicle sample calibration frame to the boundary of the frame;

Represents the predicted value of the vehicle sample category predicted by the jth anchor point in the ith grid on the lth network. Similarly,

Represents the calibration state of the vehicle sample category,

represents a sample of motor vehicles for prediction,

Indicates whether to predict the street background sample, the specific calculation is as follows:

The parameter α∈(0,1); iou _j represents the overlap rate of anchor point m _j in the i-th grid of the anchor point frame and the vehicle sample calibration frame, and miou represents the maximum overlap rate.

Step 4: Based on the loss function of the one-stage target detection algorithm model in Step 3, use the training set to train the model by gradient descent until the model converges. In the model testing stage, set the alarm time as timer. When the system model detects a motor vehicle, it automatically records the detailed category and location information it belongs to, and starts timing. After the given time timer is exceeded, if the motor vehicle is detected again If the detailed category and location information is consistent with previously detected information, an alert is issued.

The invention has the advantages that the parameter self-adaptability of the illegally parked vehicle monitoring model can be improved, and the accuracy of illegally parked vehicle monitoring can be greatly improved.

Description of drawings

FIG. 1 is a network structure diagram of the convolutional neural network of the present invention.

FIG. 2 is a structural diagram of the loss function in the convolutional neural network of the present invention.

FIG. 3 is a flow chart of the deployment of a motor vehicle parking violation detection algorithm based on a convolutional neural network of the present invention.

Detailed ways

In order to better illustrate the technical solutions of the present invention, the present invention will be further described below with reference to the accompanying drawings.

The vehicle parking violation monitoring method includes the following steps:

Step 1: Collect a large number of high-altitude motor vehicle image data, construct a motor vehicle sample data set M with a number of 10,000, a training data set T with a number of 8,000, a verification data set V with a number of 2,000, and label the number of motor vehicle categories C The value is 5, which are sports cars, off-road vehicles, large trucks, vans and ordinary cars. The training data batch size is 4, the training batch number batches is 1000, the learning rate l_rate is 0.001, and the training data is 0.001. The scale coefficient ζ between the set T and the validation data set V is 0.25, the height, width and number of channels of all images are set the same, the height h _k and width w _k of the image are respectively 416 and 416, The number r takes the value 3.

默认为1，卷积步长

默认为1，卷积神经元的激励函数f默认为leakly_relu激励函数；锚点在每一层网络中都共享，锚点集合M取值为{(10,13),(30,61),(156,198)}，即在每一层网络层中，锚点总数Λ取值为3；网络输出层采用全连接方式，其卷积核集合A取值为{(1,1,30),(1,1,30),(1,1,30)}，即输出层节点总数Ξ取值为3。Step 2: Determine the first-stage target detection model as Yolov3, and set the depth L of the convolutional neural network to 139. Among them, the height, width and dimension of the convolution kernel are set as shown in Figure 1, and the filling size of the convolution kernel is shown in Figure 1.

Default is 1, convolution stride

The default is 1, the activation function f of the convolutional neuron defaults to the leakly_relu activation function; the anchor points are shared in each layer of the network, and the anchor point set M is {(10,13),(30,61),( 156,198)}, that is, in each network layer, the total number of anchor points Λ is 3; the network output layer adopts the full connection method, and its convolution kernel set A is {(1,1,30), (1 ,1,30),(1,1,30)}, that is, the total number of output layer nodes Ξ is 3.

Step 3: As shown in Figure 2, build a parameter adaptive focus loss function LOSS, the parameter α is 0.25, and the parameter λ is 0.5.

Step 4: Based on the loss function of the one-stage target detection algorithm model in Step 3, use the training set to train the model by gradient descent until the model converges. Referring to Figure 3, real-time detection is performed using the video stream of the camera installed on the street. The alarm time timer is 3 minutes. When the system model detects a motor vehicle, it automatically records the detailed category and location information to which it belongs, and starts timing. After 3 minutes, if the detailed category and location information of the motor vehicle detected again is consistent with the previous one, an alarm will be issued to realize the management of illegal parking of motor vehicles in the street.

The content described in the embodiments of the present specification is only an enumeration of the realization forms of the inventive concept, and the protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments, and the protection scope of the present invention also extends to those skilled in the art. Equivalent technical means that can be conceived by a person based on the inventive concept.

Claims (1)

1. The method for monitoring the illegal parking of the motor vehicle comprises the following steps:

step 1: constructing a motor vehicle sample data set M, a training data set T, a verification data set V, labeling the category number C of the motor vehicle sample, the batch size batch of the training data, the number batches of the training data, the learning rate l _ rate, and a proportionality coefficient zeta between the training data set T and the verification data set V;

ζ＝Card(V)/Card(T)

wherein V ∪ T is M, C ∈ N⁺，ζ∈(0,1)，batches∈N⁺,l_rate∈N⁺，batch∈N⁺，

Representing the height and width of the image, and r represents the number of channels of the image;

step 2: determining a stage target detection model to be trained, setting the depth of a convolutional neural network as L, setting a network convolutional layer convolutional kernel set G, setting a network output layer in a full-connection mode, setting a convolutional kernel set A and a network characteristic diagram set U,

representing the kth characteristic diagram in the l-th network

The corresponding grid number and anchor point set M are specifically defined as follows:

wherein:

respectively representing the height, width and dimension of a convolution kernel, a characteristic diagram and an anchor point corresponding to the l-th network;

indicating the fill size of the layer l network convolution kernel,

representing the convolution step size of the layer I network, f representing the excitation function of the convolution neuron, theta representing the selected input feature, Λ∈ N⁺Denotes the total number of anchor points xi ∈ N in the layer I network⁺Representing the total number of output layer nodes, Φ ∈ N⁺Indicates the total number of layer I network feature maps, Δ ∈ N⁺Represents the total number of the l layer convolution kernels;

step 3: designing a parameter adaptive focus loss function, which specifically comprises the following steps:

wherein:

indicating that the jth anchor point in the ith grid on the ith network is in the image t_kThe motor vehicle sample and the street background sample confidence loss function; in the same way, the method for preparing the composite material,

a loss function representing a sample prediction box for the vehicle,

a loss function representing the motor vehicle class, λ ∈ Q being the loss function

A parameter;

and

the loss functions of the motor vehicle sample object and the street background object are respectively expressed as follows:

the probability value of the foreground motor vehicle sample predicted by the jth anchor point in the ith grid on the ith network is represented, and similarly,

representing a corresponding street context probability value;

respectively representing the abscissa and the ordinate of the central point of the prediction frame of the jth anchor point in the ith grid on the ith network, and the like

Respectively representing the abscissa and the ordinate of the central point of the motor vehicle sample calibration frame;

respectively representing the shortest Euclidean distance from the central point of the prediction frame of the jth anchor point in the ith grid on the ith network to the boundary of the frame, and the same way

Respectively representing the shortest Euclidean distance from the central point of the motor vehicle sample calibration frame to the frame boundary;

representing the predicted motor vehicle sample category value of the jth anchor point prediction in the ith grid on the ith network; in the same way, the method for preparing the composite material,

indicating the calibration status of the sample class of motor vehicles,

a sample of the motor vehicle is represented for prediction,

whether the street background sample is predicted or not is represented, and the specific calculation is as follows:

wherein the parameters α∈ (0, 1); iou_jRepresenting anchor points m_jOverlapping rate of anchor point frame and motor vehicle sample calibration frame in ith grid, miou represents maximum overlapping rate;

step 4: performing gradient descent method training on the model by using a loss function of a stage target detection algorithm model in Step 3 until the model converges; in the system operation stage, a first-order target detection model is used for extracting a network characteristic value, an anchor point is determined based on a K-means clustering method, in the system operation stage, an alarm time is set as a timer, when the system model detects a motor vehicle, the detailed type and position information of the motor vehicle are automatically recorded, timing is started, and after the given time is exceeded, if the detailed type and position information of the motor vehicle detected again are consistent with the information detected before, an alarm is sent.

Images