Improved YOLOv7-based passenger car blind area pedestrian detection method and system

Abstract

The invention provides a passenger car blind area pedestrian detection method and system based on improved YOLOv7, and relates to the technical field of artificial intelligence technology and intelligent automobiles, images of passenger car blind area pedestrians are obtained through visual shooting equipment, the collected passenger car visual angle blind area pedestrian images and 17 key point pedestrian images are marked through a Labelimg marking tool, and the passenger car blind area pedestrian detection method and system based on the improved YOLOv7 are obtained. According to the method, a passenger car blind area pedestrian detection data set is constructed, the data set is divided into a training set and a test set, and a YOLOv7 network structure is optimized by embedding a feature extraction enhancement module FEE with hole convolution operation in a detection network Head by taking a YOLOv7 algorithm as a reference through technical innovation. The cavity convolution enlarges the receptive field of the improved model under the condition of keeping the number of parameters unchanged, ensures that the size of an output feature map is kept unchanged, and enables a YOLOv7 detection network to be capable of extracting multi-scale information of a detection target.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology and intelligent automobile technology, and more specifically, particularly relates to a method and system for detecting pedestrians in the blind spot of a bus based on improved YOLOv7. Background Art

[0002] Passenger vehicles have a large carrying capacity and high transportation efficiency, and play an important role in transportation, but accidents caused by them are also frequent. Due to the large body size, high cockpit position, limited reflection range of rearview mirrors, etc., there are many blind spots in the process of turning and going straight, which makes the driver unable to fully obtain the environmental conditions around the body. Therefore, it is easy to collide with pedestrians within the inner wheel difference range of the bus, causing traffic accidents, resulting in casualties, economic losses and adverse social impacts. Based on this, the present invention proposes a bus blind spot pedestrian detection method and system based on improved YOLOv7 to detect pedestrian targets appearing in the blind spot of the bus.

[0003] At present, the target detection algorithm based on deep learning can overcome the shortcomings of low detection accuracy, susceptibility to environmental interference and weak generalization ability. The target detection algorithm based on deep learning can be divided into two types: one is a multi-stage method, representative algorithms include R-CNN and FastR-CNN, and the other is a single-stage detection algorithm, typical algorithms are YOLO series. The multi-stage method R-CNN and FastR-CNN models cannot meet the real-time requirements due to the large amount of calculation and slow detection speed. For example, the invention patent "A heavy truck blind spot target detection method based on improved YOLOv3" with publication number CN111738056B improves the detection accuracy through the K-means clustering method, but the detection accuracy is poor in small target detection, and can only detect larger pedestrian targets on the image. YOLOv7 has a speed and accuracy in the range of 5FPS to 160FPS that exceeds all known target detectors, and is currently one of the most advanced target detectors.

[0004] The main purpose of the present invention is to target the structural characteristics of passenger vehicles, which have "long, wide and high" characteristics, and the driver is on the left side of the vehicle. The driver cannot observe the road conditions around the vehicle through the rearview mirror in time. At the same time, when the passenger vehicle turns, the inner front wheel radius is greater than the inner rear wheel radius, that is, there is an inner wheel difference. The existence of the inner wheel difference is also an important reason for the blind spot of the passenger vehicle. The larger the number of axles and wheelbase of the vehicle, the larger the inner wheel difference, and the larger the right turn blind spot. The method can detect pedestrians in the blind spot of the passenger vehicle and deploy it on the mobile embedded device Jetson Xavier nx with GPU and use the device as the core controller to build a blind spot detection system. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method and system for detecting pedestrians in the blind spot of a bus based on improved YOLOv7 to solve the above problems.

[0006] A method for detecting pedestrians in a bus blind spot based on improved YOLOv7, comprising the following steps:

[0007] S1: Obtain images of pedestrians in the blind spot of a bus through a visual shooting device, annotate the images of pedestrians and 17 key points of pedestrians in the blind spot of the bus through the Labelimg annotation tool, construct a dataset for detecting pedestrians in the blind spot of a bus, and divide the dataset into a training set and a test set;

[0008] S2: Construct a pedestrian detection network for passenger cars in blind spots based on YOLOv7, improve the structure of the YOLOv7 detection network, embed a feature extraction enhancement module FEE (Feature extraction enhancement) with a hole convolution operation in the YOLOv7 detection network Head, and improve the ELAN module to the ELAN-L module and ELAN-H module for adaptive weighted feature fusion. The improved YOLOv7 detection network includes a backbone extraction network Backbone, an enhanced feature extraction network FPN, and a detection network Head. The backbone extraction network Backbone, the enhanced feature extraction network FPN, and the detection network Head are connected in sequence. The function of the entire improved YOLOv7 network is to extract features in the backbone extraction network Backbone, enhance features in the enhanced feature extraction network FPN, and predict the prior frame in the detection network Head to detect pedestrians.

[0009] S3: Train the improved YOLOv7 passenger vehicle blind spot pedestrian detection model, input the training set images processed by S1 into the detection network for training, set the training parameters, and reach the set number of learning iterations;

[0010] S4: After the training is completed, the trained improved YOLOv7 passenger vehicle blind spot pedestrian detection network weight file is saved, and an improved YOLOv7 passenger vehicle blind spot pedestrian detection model is generated for detecting pedestrians in passenger vehicle blind spots.

[0011] Preferably, S1 comprises the following steps:

[0012] S11: Obtaining an image of a pedestrian in a blind spot of the bus by using a visual shooting device installed on the rear of the bus;

[0013] S12: Process the obtained pedestrian image in the blind spot of the bus, crop and align it into a 640*640 RGB image;

[0014] S13: Use Labelimg annotation tool to annotate the pedestrian in the blind spot of the bus and the 17 key points of the pedestrian with rectangular frames to obtain an annotation file of type txt;

[0015] S14: After labeling, the images are divided into a training set of 8,000 images of pedestrians in blind spots and a test set of 2,000 images of pedestrians in blind spots.

[0016] Preferably, S2 comprises the following steps:

[0017] S21: Input the 640*640 size image processed by S1 into the backbone extraction network Backbone. The input image will be feature extracted in the backbone extraction network. The extracted feature layer is the feature set of the input image. The feature layer obtained by the backbone network is called the effective feature layer.

[0018] S22: The four effective feature layers obtained in the backbone extraction network Backbone will be fused in the enhanced feature extraction network FPN. The purpose of the fusion is to combine feature information of different scales. The effective feature layers obtained in the enhanced feature extraction network FPN are used to continue extracting features.

[0019] S23: Head is the classifier of YOLOv7. It has obtained four enhanced effective feature layers through the backbone extraction network Backbone and the enhanced feature extraction network FPN. In the detection network Head, the enhanced feature extraction network FPN continues to output four layers of feature maps of different sizes. The feature map size output by the detection head is downsampled by the pooling function in the MultiCat module and the SPPCSPC module to obtain the feature map size of 80*80, 40*40, 20*20, and 10*10.

[0020] Preferably, in S21, the backbone extraction Backbone network includes a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a first ELAN-L module, a second ELAN-L module, a third ELAN-L module, a fourth ELAN-L module, and a fifth ELAN-L module.

[0021] Preferably, the first CBS module, the second CBS module, the third CBS module, the fourth CBS module and the fifth CBS module are all CBS modules, and the first ELAN-L module, the second ELAN-L module, the third ELAN-L module, the fourth ELAN-L module and the fifth ELAN-L module are all ELAN-L modules.

[0022] Preferably, in S22, the enhanced feature extraction network FPN includes an SPPCSPC module, a first MultiCat module, a second MultiCat module, a third MultiCat module, a first ELAN-H module, a second ELAN-H module, a third ELAN-H module, a fourth ELAN-H module, a fifth ELAN-H module, a sixth ELAN-H module, a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a seventh CBS module, an eighth CBS module, and a ninth CBS module.

[0023] Preferably, the first MultiCat module, the second MultiCat module, and the third MultiCat module are all MultiCat modules, the first CBS module, the second CBS module, the third CBS module, the fourth CBS module, the fifth CBS module, the sixth CBS module, the seventh CBS module, the eighth CBS module, and the ninth CBS module are all CBS modules, and the first ELAN-H module, the second ELAN-H module, the third ELAN-H module, the fourth ELAN-H module, the fifth ELAN-H module, and the sixth ELAN-H module are all ELAN-H modules;

[0024] The MultiCat module in the FPN network will enlarge the feature map of size L through the maximum pooling operation and the average pooling operation, and then reduce and superimpose it. Finally, the feature map of size L, the feature map of size S and the feature map of size M are cat-concatenated. The sizes of S, M and L in the first MultiCat module are 20*20, 40*40 and 80*80 respectively, the sizes of S, M and L in the second MultiCat module are 40*40, 80*80 and 160*160 respectively, and the sizes of S, M and L in the third MultiCat module are 80*80, 160*160 and 320*320 respectively.

[0025] Preferably, in S23, the detection network Head includes a first FEE module, a second FEE module, a third FEE module, a fourth FEE module, a first CBS module, a second CBS module, a third CBS module, and a fourth CBS module;

[0026] Among them, the first FEE module, the second FEE module, the third FEE module and the fourth FEE module are all FEE modules, and the first CBS module, the second CBS module, the third CBS module and the fourth CBS module are all CBS modules.

[0027] Preferably, S2 is based on an improved YOLOv7 detection network and includes the following steps:

[0028] Step 1: Define the three modules ELAN-L, ELAN-H, and FEE in the cmmon.py file in the YOLOv7 detection network;

[0029] Step 2: Use the dilated convolution in the FEE module in the YOLOv7.yaml file to expand the receptive field of the YOLOv7 detection network. The receptive field is the size of the area mapped on the original image by the pixels on the feature map output by each layer of the convolutional neural network. In the dilated convolution, the size of the receptive field is controlled by the dilated coefficient (rate). That is, the larger the dilated coefficient, the larger the receptive field. The dilated convolution enables the YOLOv7 detection network to extract multi-scale information of the detection target from a more global perspective.

[0030] Step 3: Use weighted features to fuse the ELAN-L and ELAN-H modules, so that the improved network can adaptively and autonomously select good features and assign them large weights W, while assigning bad features small weights N.

[0031] Step 4: Use the parse_model() function in the yolo.py file in the YOLOv7 algorithm to parse the three modules: ELAN-L module, ELAN-H module, and FEE module.

[0032] Another technical problem to be solved by the present invention is to provide a passenger vehicle blind spot pedestrian detection system based on improved YOLOv7, including a visualization shooting device module, a passenger vehicle blind spot pedestrian detection data set module and an improved YOLOv7 passenger vehicle blind spot pedestrian detection model module.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] In the present invention, through technical innovation, the YOLOv7 algorithm is used as a benchmark, and the feature extraction enhancement module FEE with a hole convolution operation is embedded in the detection network Head to optimize the YOLOv7 network structure. The hole convolution operation is added in the feature extraction process. The hole convolution can expand the receptive field of the improved model while keeping the number of parameters unchanged. At the same time, it can ensure that the size of the output feature map remains unchanged. The addition of the hole convolution operation allows the YOLOv7 detection network to extract multi-scale information of the detection target from a more global perspective, and can systematically aggregate multiple proportional link information without losing resolution. The adaptive weighted feature fusion ELAN-L and ELAN-H modules are added to optimize the YOLOv7 network structure. The adaptive weighted feature fusion ELAN-L module and the ELAN-H module allow the improved network to adaptively and autonomously select excellent features that are important to the current task and assign large weights W, while assigning small weights N to poor features that are not important to the current task. The overall performance of the model is optimized through improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the training network model of the present invention;

[0036] Figure 2 It is a structural diagram of the network model of the present invention;

[0037] Figure 3 It is a structural diagram of the FEE feature extraction enhancement module of the present invention;

[0038] Figure 4 It is a structural diagram of the ELAN-L module and the ELAN-H module of the adaptive weighted feature fusion of the present invention;

[0039] Figure 5 It is the structural diagram of the MultiCat module of the present invention;

[0040] Figure 6 It is the structural diagram of the SPPCSPC module of the present invention;

[0041] Figure 7 is a training result diagram after the improvement of the network model of the present invention;

[0042] Figure 8 It is a PR curve diagram after the improvement of the network model of the present invention;

[0043] Fig. 9 It is a schematic diagram of the shooting position of the depth camera of the present invention;

[0044] Fig.10 Schematic diagram of the receptive field under different cavity coefficients of the present invention;

[0045] Fig.11It is a schematic diagram of improved YOLOv7 detection in an actual scenario according to an embodiment of the present invention;

[0046] Fig.12 It is a combination of the development platform of the present invention and the Ubuntu system;

[0047] Fig.13 This is the real-time detection effect of dynamic video of the system of the present invention. DETAILED DESCRIPTION

[0048] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0049] See also Figure 1-Figure 13 The present invention provides a method for detecting pedestrians in a blind spot of a bus based on an improved YOLOv7, comprising the following steps:

[0050] Step 1: Use Intel's depth camera to obtain images of pedestrians in the blind spot of the bus. The depth camera device is located at the right rear of the bus. Fig. 9 As shown, looking from back to front, the left camera is in the same position, the pedestrian walks along the x and y directions, and the installation information of the depth camera equipment is height 2.1m, HFOV 101°, VFOV 56°, yaw 20°, and pitch 10°.

[0051] Step 2: Resize the image of pedestrians in the blind spot of the bus in JPG format with a size of 1280*720 to 640*640 using Photoshop software.

[0052] Step 3: Use the image annotation software Labelimg to mark the positions of pedestrians in the blind spot pedestrian images taken by the camera with rectangular frames, and divide the txt format data set obtained after annotation into a training set and a validation set in a ratio of 8:2 to obtain the blind spot pedestrian data set.

[0053] Step 4: Construct a pedestrian detection model for passenger vehicle blind spots based on the improved YOLOv7 algorithm, improve the structure of the YOLOv7 detection network, embed a feature extraction enhancement module FEE with a hole convolution operation in the YOLOv7 detection network Head, and improve the ELAN module to an ELAN-L module and an ELAN-H module with adaptive weighted feature fusion. The improved YOLOv7 detection network includes a backbone extraction network Backbone, an enhanced feature extraction network FPN, and a detection network Head. The backbone extraction network Backbone, the enhanced feature extraction network FPN, and the detection network Head are connected in sequence, and the detection network Head outputs four layers of feature maps with sizes of 10*10, 20*20, 40*40, and 80*80. The specific steps are as follows:

[0054] First, write the FEE module, ELAN-L module, and ELAN-H module code into the cmmon.py file of the YOLOv7 model to define the three modules ELAN-L, ELAN-H, and FEE. Then use the three modules FEE module, ELAN-L module, and ELAN-H module in the YOLOv7.yaml file. Finally, add the ELAN-L module, ELAN-H module, and FEE module to the parse_model() function in the yolo.py file in the YOLOv7 detection network to parse these three modules.

[0055] Using weighted feature fusion ELAN-L and ELAN-H modules, the improved network can adaptively select good features and assign them large weights W, while assigning bad features small weights N, such as Figure 4 As shown in the figure: W1, W2, W3, and W4 weight parameters are added to the ELAN-L module, and N1, N2, N3, N4, N5, and N6 weight parameters are added to the ELAN-H module. These weight parameters are allowed to be changed during the training process. The values ​​of the parameters are automatically obtained through the training data set during the network training process. During the network training process, these weight parameters will be assigned different values. The larger the value, the more information is obtained from this feature, and vice versa. The smaller the value, the less information is obtained from this feature.

[0056] The FEE module optimization is to use the dilated convolution operation to expand the receptive field of the YOLOv7 detection network. The receptive field is the size of the area mapped on the original image by the pixel points on the feature map output by each layer of the convolutional neural network. In the dilated convolution, the size of the receptive field is controlled by the dilated coefficient (rate). That is, the larger the dilated coefficient, the larger the receptive field. The dilated convolution enables the YOLOv7 detection network to extract multi-scale information of the detection target from a more global perspective.

[0057] The dilated convolution calculates the receptive field of the convolution according to the following formula. The receptive field under different dilation coefficients is as follows: Fig.10 ;

[0058]

[0059] In the above formula, r n represents the receptive field of this layer, r n-1 represents the receptive field of the previous layer, s i represents the step size of the i-th convolution or pooling layer, and k is the convolution kernel size;

[0060] The dilated convolution uses the following formula to calculate the actual convolution kernel size after adding the dilated convolution;

[0061] K=k+(k-1)(r-1)

[0062] In the above formula, k is the original convolution kernel size, and r is the dilated coefficient rate of the dilated convolution;

[0063] The relationship between the input and output feature maps after adding the dilated convolution is shown in the formula:

[0064]

[0065] Where W is the size of the output feature map, i n is the size of the input feature map, P is the padding, k is the convolution kernel size, s is the step size, and r is the dilation coefficient rate of the dilated convolution.

[0066] Step 5: In step 4, the improved YOLOv7 algorithm is used to build a pedestrian detection model for passenger vehicle blind spots. The specific training conditions and evaluation indicators are as follows:

[0067] Based on the PyTorch learning framework, the training set images are uniformly sized to 640*640, and the training parameters are set. The training batch size is set to Batch=20, the momentum is Momentum=0.95, the weight decay parameter is 0.0005, the initial learning rate is set to Ir=0.001, and the number of training iterations is Epoch=150;

[0068] Common evaluation indicators include precision P (Precision), frame rate FPS, recall rate, precision-recall rate (PR) curve, mAP_0.5, etc. Because the present invention detects only one category, namely pedestrians, mAP = Ap;

[0069] In the formula of precision P, TP refers to the number of samples predicted correctly, and FP refers to the number of samples predicted incorrectly;

[0070]

[0071] FPS frame rate refers to the number of images that the network can predict per second. The larger the FPS, the faster the network's reasoning speed. In the formula, A represents the pre-process image preprocessing time (unit: ms), B represents the inference reasoning (unit: ms), and C represents the NMS non-maximum suppression processing time (unit: ms);

[0072]

[0073] Recall is used to measure whether a classifier can find all positive samples, that is, to evaluate the completeness of the network model prediction;

[0074]

[0075] The PR function is similar to an inverse proportional function, that is, Recall and Precision increase and decrease. The better the classifier performance, the more the PR curve protrudes to the upper right, and the larger the area enclosed by the coordinate axis. The improved YOLOv7 precision-recall curve is shown in Figure 8 As shown;

[0076] Average Precision (AP) is the area enclosed by the curve drawn by Precision and Recall and the coordinate axis. mAP is the average of AP values ​​of multiple categories. The number after _ represents the threshold of IoU. mAP_0.5 means that when IoU is set to 0.5, the AP of all pictures of each category is calculated, and then the average of all categories is calculated, that is, mAP. Because the present invention detects only one category, that is, pedestrians, mAP = Ap.

[0077] Step 6: After training, the mAP_0.5 of the improved model increased by 1.4%, the precision increased by 2.54%, and the recall rate increased by 1.53%. The trained improved YOLOv7 passenger car blind spot pedestrian detection network weight file was saved, and the model size best.pt was reduced by 14.2% compared with before improvement;

[0078] Step 7: Set up the environment for Jetson Xavier for model inference. The configuration is completed by installing Jetpack version 4.6, which includes the installation content CUDA, cuDNN, OpenCV and other components, such as Fig.12 To build the overall system after power-on;

[0079] Step 8: After assembling and powering on, configure the environment required for model reasoning on the built system platform. The installation environment includes frameworks such as Pytorch, Cython, pillow, matplotlib, and seaborn.

[0080] Step 9: Deploy the trained improved model to Jetson Xavier for real-time detection.

[0081] A passenger vehicle blind spot pedestrian detection system based on improved YOLOv7 comprises a visualization shooting device module, a passenger vehicle blind spot pedestrian detection data set module and an improved YOLOv7 passenger vehicle blind spot pedestrian detection model module.

[0082] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A method for detecting pedestrians in the blind spot of a bus based on improved YOLOv7, characterized by: The following steps are involved: S1: Obtain images of pedestrians in the blind spot of a bus through a visual shooting device, annotate the images of pedestrians and 17 key points of pedestrians in the blind spot of the bus through the Label img annotation tool, construct a dataset for detecting pedestrians in the blind spot of a bus, and divide the dataset into a training set and a test set; S2: Construct a bus blind spot pedestrian detection network based on YOLOv7, improve the structure of the YOLOv7 detection network, embed a feature extraction enhancement module FEE with a hole convolution operation in the YOLOv7 detection network Head, and improve the ELAN module to the ELAN-L module and ELAN-H module with adaptive weighted feature fusion. The improved YOLOv7 detection network includes a backbone extraction network Backbone, an enhanced feature extraction network FPN, and a detection network Head. The backbone extraction network Backbone, the enhanced feature extraction network FPN, and the detection network Head are connected in sequence. The function of the entire improved YOLOv7 network is to extract features in the backbone extraction network Backbone, enhance features in the enhanced feature extraction network FPN, and predict the prior frame in the detection network Head to detect pedestrians. S3: Train the improved YOLOv7 passenger vehicle blind spot pedestrian detection model, input the training set images processed by S1 into the detection network for training, set the training parameters, and reach the set number of learning iterations; S4: After the training is completed, the trained improved YOLOv7 passenger vehicle blind spot pedestrian detection network weight file is saved, and an improved YOLOv7 passenger vehicle blind spot pedestrian detection model is generated for detecting pedestrians in passenger vehicle blind spots.

2. The method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 1, characterized in that: S1 includes the following steps: S11: Obtaining an image of a pedestrian in a blind spot of the bus by using a visual shooting device installed on the rear of the bus; S12: Process the obtained pedestrian image in the blind spot of the bus, crop and align it into a 640*640 RGB image; S13: Label img annotation tool is used to annotate the pedestrian in the blind spot of the bus and the 17 key points of the pedestrian with rectangular frames to obtain an annotation file of type txt; S14: After labeling, the images are divided into a training set of 8,000 images of pedestrians in blind spots and a test set of 2,000 images of pedestrians in blind spots.

3. The method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 1, characterized in that: S2 includes the following steps: S21: Input the 640*640 size image processed by S1 into the backbone extraction network Backbone. The input image will be feature extracted in the backbone extraction network. The extracted feature layer is the feature set of the input image. The feature layer obtained by the backbone network is called the effective feature layer. S22: The four effective feature layers obtained in the backbone extraction network Backbone will be fused in the enhanced feature extraction network FPN. The purpose of the fusion is to combine feature information of different scales. The effective feature layers obtained in the enhanced feature extraction network FPN are used to continue extracting features. S23: Head is the classifier of YOLOv7. Four enhanced effective feature layers have been obtained through the backbone extraction network Backbone and the enhanced feature extraction network FPN. In the detection network Head, four layers of feature maps of different sizes are continued to be output through the enhanced feature extraction network FPN. The size of the feature map output by the detection head is downsampled by the pooling function in the MultiCat module and the SPPCSPC module to obtain the feature map size of 80*80, 40*40, 20*20, and 10*10.

4. A method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 3, characterized in that: In S21, the backbone extraction Backbone network includes a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a first ELAN-L module, a second ELAN-L module, a third ELAN-L module, a fourth ELAN-L module, and a fifth ELAN-L module.

5. A method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 4, characterized in that: The first CBS module, the second CBS module, the third CBS module, the fourth CBS module and the fifth CBS module are all CBS modules, and the first ELAN-L module, the second ELAN-L module, the third ELAN-L module, the fourth ELAN-L module and the fifth ELAN-L module are all ELAN-L modules.

6. A method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 3, characterized in that: In S22, the enhanced feature extraction network FPN includes an SPPCSPC module, a first MultiCat module, a second MultiCat module, a third MultiCat module, a first ELAN-H module, a second ELAN-H module, a third ELAN-H module, a fourth ELAN-H module, a fifth ELAN-H module, a sixth ELAN-H module, a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a fifth CBS module, a sixth CBS module, a seventh CBS module, an eighth CBS module, and a ninth CBS module.

7. A method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 6, characterized in that: The first MultiCat module, the second MultiCat module, and the third MultiCat module are all MultiCat modules, the first CBS module, the second CBS module, the third CBS module, the fourth CBS module, the fifth CBS module, the sixth CBS module, the seventh CBS module, the eighth CBS module, and the ninth CBS module are all CBS modules, and the first ELAN-H module, the second ELAN-H module, the third ELAN-H module, the fourth ELAN-H module, the fifth ELAN-H module, and the sixth ELAN-H module are all ELAN-H modules; The MultiCat module in the FPN network will enlarge the feature map of size L through the maximum pooling operation and the average pooling operation, and then reduce and superimpose it. Finally, the feature map of size L, the feature map of size S and the feature map of size M are cat-concatenated. The sizes of S, M and L in the first MultiCat module are 20*20, 40*40 and 80*80 respectively, the sizes of S, M and L in the second MultiCat module are 40*40, 80*80 and 160*160 respectively, and the sizes of S, M and L in the third MultiCat module are 80*80, 160*160 and 320*320 respectively.

8. The method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 3, characterized in that: In S23, the detection network Head includes a first FEE module, a second FEE module, a third FEE module, a fourth FEE module, a first CBS module, a second CBS module, a third CBS module, and a fourth CBS module; Among them, the first FEE module, the second FEE module, the third FEE module and the fourth FEE module are all FEE modules, and the first CBS module, the second CBS module, the third CBS module and the fourth CBS module are all CBS modules.

9. The method for detecting pedestrians in a passenger car blind spot based on improved YOLOv7 as claimed in claim 3, characterized in that: S2 is based on the improved YOLOv7 detection network, including the following steps: Step 1: Define the three modules ELAN-L, ELAN-H, and FEE in the cmmon.py file in the YOLOv7 detection network; Step 2: Use the dilated convolution in the FEE module in the YOLOv7.yaml file to expand the receptive field of the YOLOv7 detection network; Step 3: Use weighted features to fuse the ELAN-L and ELAN-H modules, so that the improved network can adaptively and autonomously select good features and assign them large weights W, while assigning bad features small weights N. Step 4: Use the parse_model() function in the yolo.py file in the YOLOv7 algorithm to parse the three modules: ELAN-L module, ELAN-H module, and FEE module.

10. A passenger car blind spot pedestrian detection system based on improved YOLOv7, characterized in that: It includes a visualization shooting equipment module, a passenger vehicle blind spot pedestrian detection dataset module, and an improved YOLOv7 passenger vehicle blind spot pedestrian detection model module.

Images