CN113989495A - Vision-based pedestrian calling behavior identification method - Google Patents

Abstract

The invention discloses a vision-based pedestrian calling behavior identification method, which comprises the following steps of: image preprocessing and intent inference. The invention adopts a computer vision method to accurately and efficiently identify the pedestrians with the taxi calling behavior from the image, realizes that the automatic taxi driving finds the passengers more efficiently, improves the use efficiency of the automatic taxi driving, and also improves the trip efficiency of the passengers. The invention adopts the spatial reasoning network to realize the reasoning of the pedestrian car-calling behavior, reduces the dependence on time dimension information, reduces the time characteristic extraction process compared with the traditional behavior recognition algorithm, can simplify the network and improve the real-time performance of behavior reasoning. The invention adopts a set of fusion rules with logical interpretability to realize the fusion of random forests and graph convolution networks, the characteristic of logical interpretability can improve the environmental adaptability and the behavior recognition precision of the algorithm, and the fusion algorithm can realize more stable and accurate reasoning on the pedestrian car-calling intention.

Description

A Vision-Based Pedestrian Car-hailing Behavior Recognition Method

technical field

The invention belongs to the field of vehicle intelligence, and in particular relates to a method for recognizing pedestrian behavior intention by an automatic driving taxi.

Background technique

The behavior of vehicles in traffic scenes to identify pedestrians belongs to the category of vehicle intelligence. Accurate and effective identification of pedestrians' car-hailing intentions can help autonomous taxis quickly find pedestrians with car-hailing intentions on the road, which is important for improving pedestrians' travel efficiency, improving the use efficiency of autonomous taxis, and avoiding traffic congestion. significance.

Pedestrian car-hailing behavior recognition refers to the use of computer vision methods to analyze pedestrians in traffic scenes to find pedestrians with the intention of car-hailing. Traffic scenes are highly complex, and the number and types of traffic participants (including pedestrians, vehicles, cyclists, etc.) are much higher than other application scenarios, which increases the difficulty of behavior recognition. Compared with other behaviors of pedestrians (walking, running, cycling, etc.), the car-hailing behavior has obvious randomness and transient characteristics: First, any pedestrian in the current scene may be transformed into a car-hailing vehicle at any time. In addition, the car-hailing behavior has obvious instantaneous characteristics, and the driver only needs a single image to determine whether a person has the car-hailing intention, and does not need to consider the consecutive frames before and after the image. Information. Based on the above two characteristics, traditional behavior recognition algorithms based on 3DCNN (3D Convolutional Neural Network) and LSTM (Long Short Term Memory Network) cannot be applied to the reasoning of car-hailing intent with instantaneous characteristics. Pedestrian gestures are the key information to express the intention of pedestrians, and most of the current gesture recognition algorithms are mainly used in indoor scenes, and vision-based gesture recognition algorithms have high requirements on the resolution of the hand contour in the image, but smart cars are equipped with The in-vehicle cameras of 100% cannot achieve such high-quality images in complex traffic scenes.

SUMMARY OF THE INVENTION

In order to solve the above-mentioned problems existing in the prior art, the present invention aims to design a pedestrian car-calling behavior recognition method with strong environmental adaptability, high recognition accuracy and based on vision, which can process the images collected by the vehicle-mounted camera, and real-timely identify the calls in the images in real time. Accurate identification of vehicles and pedestrians to help autonomous taxis find passengers more efficiently.

In order to achieve the above object, the technical solution of the present invention is as follows: a vision-based pedestrian car-hailing behavior recognition method, comprising the following steps:

A. Image preprocessing

The target detection algorithm and the human body key point extraction algorithm are used to preprocess the image, and the pedestrian detection frame D and the pedestrian key point parameter K corresponding to each detection frame are obtained. Facial attention is a key clue to determine whether it has the intention of hailing a car. In a real scene, in the process of pedestrian hailing a car, pedestrians will have a high degree of attention to taxis. The reasoning of facial attention is carried out from two aspects. First, the facial key points detected in the human body key point detection are used for inference, and the difference h _p between the abscissas of the left ear key point and the right ear key point is used as the benchmark. Taking σ as the magnification factor, a square frame S with side length σh _p is formed as the face area; when the lateral distance h _f between the left ear key point and the nose key point is greater than h _p , it means that the pedestrian's face is at a positive angle from the opposite side. For taxis, that is, pedestrians pay less attention to vehicles; when h _f is less than h _p , input the face region S into the facial attention depth network to calculate the pedestrian's facial attention probability; the facial attention depth network includes the front network and the rear network, the front network is a feature extraction network, and Resnet50 is used as the benchmark network to extract facial features; the rear network is a feature connection network composed of fully connected layers, and the facial features extracted by the front network are connected. global feature, the output is the facial attention probability ρ _f ;

B. Intentional reasoning

Using the combination of random forest algorithm and graph convolutional network to infer pedestrian intent, the specific steps are as follows:

B1. The random forest algorithm is used to infer the relationship between the connection angle between the key points of the human body and the intention of pedestrians. The input of the random forest is the connection angle of the key points of the human body. The strong key point angle is used as the input of random forest, including the connection angle of neck key point, left shoulder key point, right shoulder key point, left elbow key point, and right elbow key point as vertices. The output of random forest is that pedestrians have car-hailing Probability ρ _r of intent.

B2. The graph convolution network is used to infer the relationship between the position of the human body key points and the pedestrian intent. The input of the graph convolution network is the human body graph model G(v,e), where v is the node of the human body graph model, that is, the human body key point, The node features are the coordinates of the key points, and e is the edge of the human body graph model, that is, the connection between the nodes. Since the size of the detection frame D obtained by the target detection is not fixed, in order to reduce the influence of the size of the detection frame on the intention inference, coordinate transformation is used to convert the image coordinates of the human body key points into the associated coordinates with the human neck key point as the origin:

Among them, x _inew and y _inew are the abscissa and ordinate after the transformation of the key point of the ith person; u _i and v _i are the abscissa and ordinate before the transformation of the key point of the ith person; u ₁ and v ₁ are the abscissa and ordinate of the neck key point.

The process of graph convolutional network is:

A是人体图模型的邻接矩阵；

是人体图模型的度矩阵；H^(l)是第l层图卷积的输出特征，H^(l+1)为第l+1层图卷积的输出特征；W^(l)为第l层图卷积的参数矩阵；

是激活函数；Z是图卷积网络的输出，即行人具有召车意图的概率ρ_g；H^(z)是最后一层图卷积的特征矩阵；W^(z)是最后一层图卷积的参数矩阵；readout(·)是由全连接层组成的图读出网络，实现将人体图模型中的所有节点特征聚合连接。in,

A is the adjacency matrix of the human figure model;

is the degree matrix of the human graph model; H ^(l) is the output feature of the lth layer graph convolution, H ^(l+1) is the output feature of the l+1th layer graph convolution; W ^(l) is the lth layer The parameter matrix of graph convolution;

is the activation function; Z is the output of the graph convolution network, that is, the probability ρ _g that the pedestrian has the intention to call a car; H ^(z) is the feature matrix of the last layer of graph convolution; W ^(z) is the last layer of graph convolution The parameter matrix of ; readout( ) is a graph readout network composed of fully connected layers, which realizes the aggregation and connection of all node features in the human graph model.

B3. Algorithm fusion

Through random forest and graph convolutional network, the probability of the pedestrian's intention to call a car is obtained. The random forest output probability ρ _r and the graph convolution network output probability ρ _g are respectively obtained. In order to obtain more stable and accurate intention reasoning, a set of logically feasible The explained fusion rule realizes the fusion of random forest and graph convolutional network. The fusion rule is as follows:

当p_g＞0.5且p_r＜0.5时，则意味着随机森林算法和图卷积网络算法具有不同的推理结果，图卷积网络的推理结果为行人具有召车意图，随机森林的推理结果为行人没有召车意图，为了得到一个更准确的推理结果，面部注意力概率p_f作为动态权重对p_g和p_r实现动态加权平均，即，当p_f＞0.5，意味着行人具有较高的召车概率，则赋予图卷积网络的输出一个更高的权重，而随机森林的输出赋予一个较低的权重；当p_f＜0.5时，则赋予随机森林的输出一个更高的权重，而赋予图卷积网络的输出一个更低的权重；当p_g＜0.5且p_r＞0.5时，则意味着另一种随机森林算法和图卷积网络算法具有不同的推理结果的情况，图卷积网络的推理结果为行人没有召车意图，而随机森林的推理结果为行人具有召车意图，当p_f＞0.5时，意味着随机森林的推理结果有更高的概率为正确的结果，则随机森林的输出赋予更高的权重，而图卷积网络的输出赋予更低的权重；反之，当p_f＜0.5时，则图卷积网络的输出赋予更高的权重，而随机森林的输出赋予更低的权重。Among them, p is the probability that the pedestrian has the intention to call a car after fusion. When p _g >0.5 and p _r >0.5 or p _g <0.5 and p _r <0.5, it means that the random forest algorithm and the graph convolutional network algorithm have the same inference result, then the fusion probability p is

When p _g > 0.5 and p _r < 0.5, it means that the random forest algorithm and the graph convolutional network algorithm have different inference results. The inference result of the graph convolution network is that the pedestrian has the intention to call a car, and the inference result of the random forest is Pedestrians have no intention to call a car. In order to obtain a more accurate inference result, the facial attention probability p _{f is} used as a dynamic weight to achieve a dynamic weighted average of p _g and p _r , that is, when p _f > 0.5, it means that the pedestrian has a higher car-hailing probability, the output of the graph convolutional network is given a higher weight, and the output of the random forest is given a lower weight; when p _f < 0.5, the output of the random forest is given a higher weight, and Give the output of the graph convolutional network a lower weight; when p _g < 0.5 and p _r > 0.5, it means that another random forest algorithm and the graph convolutional network algorithm have different inference results. The reasoning result of the product network is that the pedestrian has no intention to call a car, while the reasoning result of the random forest is that the pedestrian has the intention to call a car. When p _f > 0.5, it means that the reasoning result of the random forest has a higher probability of being the correct result, then The output of the random forest is given a higher weight, while the output of the graph convolutional network is given a lower weight; conversely, when p _f < 0.5, the output of the graph convolutional network is given a higher weight, and the output of the random forest is given a higher weight. assign lower weights.

Compared with the prior art, the beneficial effects and benefits of the present invention are as follows:

1. The present invention uses the method of computer vision to accurately and efficiently identify pedestrians with car-hailing behaviors from images, so that the autonomous taxi can find passengers more efficiently, improve the use efficiency of the autonomous taxi, and improve the passenger's safety. travel efficiency.

2. The present invention adopts a spatial reasoning network to realize the reasoning of pedestrian car-hailing behavior, which reduces the dependence on time dimension information. Compared with the traditional behavior recognition algorithm, the process of time feature extraction is reduced, which can simplify the network and improve the behavior. Real-time inference.

3. The present invention adopts a set of logically interpretable fusion rules to realize the fusion of random forest and graph convolutional network. The logically interpretable characteristics can improve the environmental adaptability of the algorithm and the accuracy of behavior recognition, and realize the fusion algorithm. More stable and accurate reasoning about pedestrian car-hailing intentions.

Description of drawings

FIG. 1 is a schematic flow chart of the present invention.

Figure 2 is a schematic diagram of human key points extracted by OpenPose.

Figure 3 is a schematic diagram of the facial attention deep network.

Figure 4 is a schematic diagram of a random forest.

Figure 5 is a schematic diagram of a graph convolutional network.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings. As shown in Figure 1, a visual-based pedestrian car-hailing behavior recognition method includes the following steps:

A. Image preprocessing

Using Yolov5 as the target detection method and the human body key point extraction algorithm OpenPose to realize the preprocessing of the image, the pedestrian detection frame D and the pedestrian key point parameter K corresponding to each detection frame are obtained. The parameters of the key points are shown in the figure 2, the correspondence between the sequence of key points and the body parts is:

The detection frame provided by target detection can improve the accuracy of human key point extraction. In the process of car-hailing intention reasoning, the facial attention of the human body is the key clue to determine whether it has the car-hailing intention. In the real scene, pedestrians will have a high degree of attention to the taxi during the process of hailing a car. For the reasoning of facial attention, the present invention is mainly carried out from two aspects. First, the facial key points detected in the human body key point detection are used for reasoning, and the difference h _p between the abscissas of the key point 16 and the key point 17 is used as the benchmark. , with σ=1.2 as the magnification factor, a square frame S with side length σh _p is formed as the face area. When the horizontal distance h _f between the key point 16 and the key point 0 is greater than h _p , it means that the pedestrian's face is relatively The angle of the side faces the taxi, that is, the pedestrian pays less attention to the vehicle, and the facial attention probability ρ _f = 0.1 is set. When h _f is less than h _p , it is difficult to judge whether the pedestrian pays attention to the vehicle through the above. Therefore, set the The facial area S is input into the facial attention depth network to calculate its facial attention probability. The schematic diagram of the facial attention depth network is shown in Figure 3. It mainly consists of two parts. The first part is the feature extraction network. Resnet50 is used as the benchmark network. Facial features, the latter part is a feature connection network composed of fully connected layers, and the features extracted in the former part are connected to obtain global features, and the output is the facial attention probability ρ _f .

B. Intentional reasoning

Through step A, the target detection frame D of the pedestrian, the human key points K of the pedestrian in the target detection frame and the facial attention probability ρ _f of the corresponding pedestrian can be obtained. The present invention adopts the combination of random forest algorithm and graph convolution network to perform pedestrian intention reasoning.

B1. Random forest mainly infers the relationship between the connection angle between human key points and pedestrian intent. Therefore, the input of the random forest is the connection angle of the key points of the human body. In order to prevent the phenomenon of overfitting, in the present invention, some key point angles that have a strong relationship with the pedestrian car-hailing are selected as the input of the random forest, including the key point 1. The key point 2, key point 3, key point 5 and key point 6 are the connection angles of the vertices. The output of the random forest is that the probability that the pedestrian has the intention to call a car is ρ _r , and the input key point connection angle is based on the key point. ∠318, ∠6111, ∠418, ∠7111, ∠618, ∠617 with 1 as the vertex; ∠123, ∠124 with the key point 2 as the vertex; ∠156, ∠157 with the key point 5 as the vertex; Point 3 is ∠234, ∠438, ∠134; the key point 6 is ∠567, ∠7611, ∠167.

The schematic diagram of the random forest is shown in Figure 4. The random forest is composed of N independent decision trees, where N=55. Different data sets are used to train different decision trees to obtain the corresponding model including the training parameters. Each decision tree is a specific classifier and makes independent decisions based on the input data. The process of decision aggregation adopts the majority voting method, and the output is the ratio of the number of decision trees whose decision is the car-hailing intention to the total number of decision trees, that is, the probability ρ _r that the pedestrian has the car-hailing intention.

B2. The graph convolutional network mainly infers the relationship between the position of the human body key points and the pedestrian intent. Therefore, the input of the graph convolutional network is the human body graph model G(v,e), where v is the node of the human body graph model, that is, the human body key point, the node feature is the coordinate of the key point, e is the edge of the human body graph model, that is, the connection between the nodes. Since the size of the detection frame D obtained by the target detection is not fixed, in order to reduce the influence of the size of the detection frame on the intention reasoning, coordinate transformation is used to convert the image coordinates of the key points of the human body into the associated coordinates with the key point 1 as the origin:

Among them, x _inew and y _inew are the abscissa and ordinate after the transformation of the ith person's key point, u _i and v _i are the abscissa and ordinate before the transformation of the ith person's key point; u ₁ and v ₁ is the abscissa and ordinate of key point 1.

The schematic diagram of the graph convolutional network is shown in Figure 5. The human body graph model is input into the graph convolutional network, and each node feature of the human body transfers the node feature to the adjacent nodes along the edges between the nodes, and each Nodes also aggregate the features transmitted from adjacent nodes to realize the transfer and aggregation of node features along the edge. In order to enhance the expressive ability of the model, after each layer of graph convolution, the activation function RELU is used to realize the nonlinear mapping of node features. Finally, A graph readout network composed of fully connected layers is used to aggregate and connect the features of all nodes to obtain the final classification result.

The process of graph convolutional network can be summarized as:

A是人体图模型的邻接矩阵；

是人体图模型的度矩阵；H^(l)是第l层图卷积的输出特征，H^(l+1)为第l+1层图卷积的输出特征；W^(l)为第l层图卷积的参数矩阵；

是激活函数RELU；Z是图卷积网络的输出，即行人具有召车意图的概率ρ_g；H^(z)是最后一层图卷积的特征矩阵；W^(z)是最后一层图卷积的参数矩阵；readout(·)是由全连接层组成的图读出网络，能够实现将人体图模型中的所有节点特征聚合连接。in,

A is the adjacency matrix of the human figure model;

is the degree matrix of the human graph model; H ^(l) is the output feature of the lth layer graph convolution, H ^(l+1) is the output feature of the l+1th layer graph convolution; W ^(l) is the lth layer The parameter matrix of graph convolution;

is the activation function RELU; Z is the output of the graph convolution network, that is, the probability ρ _g of the pedestrian with the intention to call a car; H ^(z) is the feature matrix of the last layer of graph convolution; W ^(z) is the last layer of graph volume The parameter matrix of the product; readout( ) is a graph readout network composed of fully connected layers, which can aggregate and connect all node features in the human graph model.

B3. Algorithm fusion

Through random forest and graph convolutional network, the probability ρ _r and ρ _g of pedestrians with car-calling intention are obtained respectively. In order to obtain more stable and accurate intention reasoning, the present invention proposes a set of logically interpretable fusion rules to realize random Fusion of forest and graph convolutional network, the fusion rules are as follows:

当p_g＞0.5且p_r＜0.5时，则意味着随机森林算法和图卷积网络算法具有不同的推理结果，图卷积网络的推理结果为行人具有召车意图，随机森林的推理结果为行人没有召车意图，为了得到一个更准确的推理结果，面部注意力p_f作为动态权重对p_g和p_r实现动态加权平均，即，当p_f＞0.5，意味着行人具有较高的召车概率，则赋予图卷积网络的输出一个更高的权重，而随机森林的输出赋予一个较低的权重；当p_f＜0.5时，则赋予随机森林的输出一个更高的权重，而赋予图卷积网络的输出一个更低的权重；当p_g＜0.5且p_r＞0.5时，则意味着另一种随机森林算法和图卷积网络算法具有不同的推理结果的情况，图卷积网络的推理结果为行人没有召车意图，而随机森林的推理结果为行人具有召车意图，当p_f＞0.5时，意味着随机森林的推理结果有更高的概率为正确的结果，则随机森林的输出赋予更高的权重，而图卷积网络的输出赋予更低的权重；反之，当p_f＜0.5时，则图卷积网络的输出赋予更高的权重，而随机森林的输出赋予更低的权重。Among them, p is the probability that the pedestrian has the intention to call a car after fusion. When p _g >0.5 and p _r >0.5 or p _g <0.5 and p _r <0.5, it means that the random forest algorithm and the graph convolutional network algorithm have the same inference result, then the fusion probability p is

When p _g > 0.5 and p _r < 0.5, it means that the random forest algorithm and the graph convolutional network algorithm have different inference results. The inference result of the graph convolution network is that the pedestrian has the intention to call a car, and the inference result of the random forest is Pedestrians have no intention to call a car. In order to obtain a more accurate inference result, the facial attention p _{f is} used as a dynamic weight to achieve a dynamic weighted average of p _g and p _r , that is, when p _f > 0.5, it means that the pedestrian has a higher calling car probability, the output of the graph convolutional network is given a higher weight, and the output of the random forest is given a lower weight; when p _f < 0.5, the output of the random forest is given a higher weight, and the output of the random forest is given a lower weight; The output of the graph convolutional network has a lower weight; when p _g <0.5 and p _r >0.5, it means that another random forest algorithm and the graph convolutional network algorithm have different inference results. The inference result of the network is that the pedestrian has no intention to call a car, while the inference result of the random forest is that the pedestrian has the intention to call a car. When p _f > 0.5, it means that the inference result of the random forest has a higher probability of being the correct result, then the random forest The output of the forest is given a higher weight, while the output of the graph convolutional network is given a lower weight; conversely, when p _f < 0.5, the output of the graph convolutional network is given a higher weight, and the output of the random forest is given a higher weight. lower weight.

The present invention is not limited to this embodiment, and any equivalent ideas or changes within the technical scope disclosed in the present invention are included in the protection scope of the present invention.

Claims (1)

1. a vision-based pedestrian car-hailing behavior recognition method, is characterized in that: comprise the following steps: A. Image preprocessing The target detection algorithm and the human body key point extraction algorithm are used to preprocess the image, and the pedestrian detection frame D and the pedestrian key point parameter K corresponding to each detection frame are obtained. Facial attention is a key clue to determine whether it has the intention of hailing a car. In a real scene, in the process of pedestrian hailing a car, pedestrians will have a high degree of attention to the taxi; the reasoning of facial attention, from two aspects First, use the facial key points detected in the human body key point detection to infer, take the difference h _p between the abscissas of the left ear key point and the right ear key point as the benchmark, and use σ as the amplification factor to form a side length of The square frame S of σh _p is used as the face area; when the lateral distance h _f between the left ear key point and the nose key point is greater than h _p , it means that the pedestrian's face is facing the taxi at the opposite side angle, that is, the pedestrian's attention to the vehicle smaller; when h _f is less than h _p , input the face region S into the facial attention depth network to calculate the pedestrian's facial attention probability; the facial attention depth network includes the front network and the back network, and the front network is the feature extraction The network uses Resnet50 as the benchmark network to extract facial features; the rear network is a feature connection network composed of fully connected layers, which realizes the connection of the facial features extracted by the front network to obtain global features, and the output is the facial attention probability ρ _f ; B. Intentional reasoning Using the combination of random forest algorithm and graph convolutional network to infer pedestrian intent, the specific steps are as follows: B1. The random forest algorithm is used to infer the relationship between the connection angle between the key points of the human body and the intention of pedestrians. The input of the random forest is the connection angle of the key points of the human body. The strong key point angle is used as the input of random forest, including the connection angle of neck key point, left shoulder key point, right shoulder key point, left elbow key point, and right elbow key point as vertices. The output of random forest is that pedestrians have car-hailing the probability of intent ρ _r ; B2. The graph convolution network is used to infer the relationship between the position of the human body key points and the pedestrian intent. The input of the graph convolution network is the human body graph model G(v,e), where v is the node of the human body graph model, that is, the human body key point, The node features are the coordinates of the key points, and e is the edge of the human body graph model, that is, the connection between the nodes; since the size of the detection frame D obtained by the target detection is not fixed, in order to reduce the influence of the size of the detection frame on the intention reasoning, the coordinates are used. The transformation realizes the transformation of the image coordinates of the key points of the human body into the associated coordinates with the key points of the neck of the human body as the origin:

Among them, x _inew and y _inew are the abscissa and ordinate after the transformation of the key point of the ith person; u _i and v _i are the abscissa and ordinate before the transformation of the key point of the ith person; u ₁ and v ₁ are the abscissa and ordinate of the neck key point; The process of graph convolutional network is:

in,

A is the adjacency matrix of the human figure model;

is the degree matrix of the human graph model; H ^(l) is the output feature of the lth layer graph convolution, H ^(l+1) is the output feature of the l+1th layer graph convolution; W ^(l) is the lth layer The parameter matrix of graph convolution;

is the activation function; Z is the output of the graph convolution network, that is, the probability ρ _g that the pedestrian has the intention to call a car; H ^(z) is the feature matrix of the last layer of graph convolution; W ^(z) is the last layer of graph convolution The parameter matrix of ; readout( ) is a graph readout network composed of fully connected layers, which realizes the aggregation and connection of all node features in the human graph model; B3. Algorithm fusion Through random forest and graph convolutional network, the probability of the pedestrian's intention to call a car is obtained. The random forest output probability ρ _r and the graph convolution network output probability ρ _g are respectively obtained. In order to obtain more stable and accurate intention reasoning, a set of logically feasible The explained fusion rule realizes the fusion of random forest and graph convolutional network. The fusion rule is as follows:

Among them, p is the probability that the pedestrian has the intention to call a car after fusion; when p _g >0.5 and p _r >0.5 or p _g <0.5 and p _r <0.5, it means that the random forest algorithm and the graph convolutional network algorithm have the same Inference result, then the fusion probability p is

When p _g > 0.5 and p _r < 0.5, it means that the random forest algorithm and the graph convolutional network algorithm have different inference results. The inference result of the graph convolution network is that the pedestrian has the intention to call a car, and the inference result of the random forest is Pedestrians have no intention to call a car. In order to obtain a more accurate inference result, the facial attention probability p _{f is} used as a dynamic weight to achieve a dynamic weighted average of p _g and p _r , that is, when p _f > 0.5, it means that the pedestrian has a higher car-hailing probability, the output of the graph convolutional network is given a higher weight, and the output of the random forest is given a lower weight; when p _f < 0.5, the output of the random forest is given a higher weight, and Give the output of the graph convolutional network a lower weight; when p _g < 0.5 and p _r > 0.5, it means that another random forest algorithm and the graph convolutional network algorithm have different inference results. The reasoning result of the product network is that the pedestrian has no intention to call a car, while the reasoning result of the random forest is that the pedestrian has the intention to call a car. When p _f > 0.5, it means that the reasoning result of the random forest has a higher probability of being the correct result, then The output of the random forest is given a higher weight, while the output of the graph convolutional network is given a lower weight; conversely, when p _f < 0.5, the output of the graph convolutional network is given a higher weight, and the output of the random forest is given a higher weight. assign lower weights.

Images