Sensing algorithm of intelligent vehicle lamp self-closed loop and auxiliary driving system

Abstract

The sensing algorithm of the intelligent vehicle lamp self-closed-loop and auxiliary driving system is mainly based on high-mount stop lamp state detection, that is, the high-mount stop lamp state of a front vehicle is detected and input to a decision regulation and control system of the vehicle and the intelligent vehicle lamp; and a YOLOv8-BL network model which is re-optimized and built based on a YOLOv8n network model is utilized to complete the detection of the vehicle high-mount stop lamp in a front view image collected by a front view camera in the intelligent auxiliary driving and intelligent vehicle lamp self-closed loop system. According to the sensing algorithm of the intelligent vehicle lamp self-closed loop and auxiliary driving system, the intelligent auxiliary driving and intelligent vehicle lamp self-closed loop system can obtain richer environment information by effectively sensing the state of the high-mount stop lamp of the front vehicle, the future state can be predicted more accurately, and the driving quality is improved. Humanized intelligent vehicle lamp and vehicle control decisions are made, sensory conflicts are reduced, and the riding comfort, the driving safety and the driving safety are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of an intelligent vehicle lamp self-closing loop and an assisted driving system, and in particular to a perception algorithm of an intelligent vehicle lamp self-closing loop and an assisted driving system. Background Art

[0002] With the continuous improvement of intelligent perception technology, it has been increasingly used in the field of intelligent connected vehicles in recent years, greatly promoting the development of intelligent automotive industry. The intelligent closed-loop system of headlights can realize automatic adjustment and intelligent response of headlights by integrating advanced algorithms and sensor technologies. For example, when the perception algorithm detects a vehicle ahead, it can automatically adjust the brightness and illumination angle of the headlights to avoid dazzling the vehicle ahead while ensuring that the driver of the vehicle has sufficient lighting vision. In addition, the perception algorithm can also automatically adjust the brightness and color temperature of the headlights according to road conditions and weather conditions to provide the best lighting effect.

[0003] In the current intelligent driving and intelligent headlight control systems, the vehicle mainly estimates the relative distance and relative speed to the surrounding vehicles to assist in making vehicle and headlight control instructions. If the relative speed or relative distance estimation is inaccurate, it will greatly affect the intelligent experience of the entire vehicle, reduce driving comfort, and seriously affect driving safety. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a perception algorithm for an intelligent self-closed-loop headlight and assisted driving system. By effectively perceiving the status of the high-mounted brake lights of the vehicle ahead, the intelligent assisted driving and the intelligent self-closed-loop headlight system can obtain richer environmental information, more accurately predict future status, make humanized intelligent headlight and vehicle control decisions, reduce sensory conflicts, and improve ride comfort, driving safety, and driving safety.

[0005] According to an embodiment of the present invention, a perception algorithm of an intelligent headlight self-closed loop and assisted driving system is mainly based on high-mounted brake light status detection, that is, detecting the status of the high-mounted brake lights of the vehicle in front and inputting it into the decision-making and control system of the vehicle and the intelligent headlights; using the YOLOv8-BL network model re-optimized and rebuilt based on the YOLOv8n network model to complete the vehicle high-mounted brake light detection in the front view image collected by the front camera in the intelligent assisted driving and intelligent headlight self-closed loop system.

[0006] The beneficial effect of the present invention is that it can enable the intelligent assisted driving and intelligent car light self-closed loop system to obtain richer environmental information, more accurately predict future states, make humanized intelligent car light and vehicle control decisions, reduce sensory conflicts, and improve ride comfort and driving safety.

[0007] According to one embodiment of the present invention, the YOLOv8n network model structure includes an input for data enhancement, a backbone network (BackBone) for image feature extraction, a bottleneck module (Neck) for feature fusion, and a decoupling head for separating classification from a detection head.

[0008] According to one embodiment of the present invention, an SE channel attention mechanism module is added to the YOLOv8n network model structure to make the YOLOv8n network model more suitable for feature extraction of key areas of the image. The SE channel attention mechanism module adaptively adjusts the importance of each channel in the feature map by learning the attention of different channels in the image.

[0009] According to one embodiment of the present invention, a lightweight convolution GSConv module and a VoV-GSCSPC module are used to replace the Cross Stage Partial Network Fusion (C2f) fusion module in the YOLOv8n network model, local and global structural features are captured by aggregating multi-layer neighborhood information, and the input feature map is grouped and processed, and each group independently performs convolution operations and parallel calculations.

[0010] According to one embodiment of the present invention, the YOLOv8n network model uses a SpatialPyr-amid Pooling Fast (SPPF) feature image pyramid pooling module in the backbone network.

[0011] According to one embodiment of the present invention, the SE channel attention mechanism module is embedded in the backbone network (BackBone) to form a new feature extraction network SE-BackBone. In the SE-BackBone network, the output feature map Q1 of the SPPF layer is passed to the SE channel attention mechanism module as an input feature, and is sequentially compressed by a global average pooling layer and excited by two fully connected layers to obtain a channel weighted feature map Q2. The feature map Q2 is multiplied by the output feature map Q1 of the SPPF layer channel by channel to recalibrate the importance of the input feature, thereby obtaining the output feature map Q3 of the SE channel attention mechanism module.

[0012] According to one embodiment of the present invention, a channel attention SE channel attention mechanism module is used to extract image features. The SE channel attention mechanism module enhances the extraction of image difference information by weighted processing of features of different channels of the image, enhances the learning of the target key feature area by the YOLOv8-BL network model, and helps to improve the detection accuracy of the YOLOv8-BL network model for the high-mounted brake light area in the image.

[0013] According to one embodiment of the present invention, the mapping relationship of the SE channel attention mechanism module is shown in formula (1), formula (2) and formula (3):

[0014]

[0015] L we (u c , e c )=e c u c (3)

[0016] The symbols in formula (1), formula (2) and formula (3) mean:

[0017] L sq Indicates compressed calculation;

[0018] L ex represents incentive calculation;

[0019] L we represents weighted calculation;

[0020] H represents the height of the input feature Q1;

[0021] W represents the width of the input feature Q1;

[0022] i and j represent the continuous values ​​of pixels. The value range of i is [0, H], and the value range of j is [0, W];

[0023] u c Represents each feature channel;

[0024] s represents the number of channels after compression calculation;

[0025] Represents the nonlinear activation function Relu;

[0026] W0 represents the total parameters of the fully connected (FC) calculation;

[0027] W1, W2, and g represent the parameters of the fully connected calculation of each layer;

[0028] θ represents the Sigmoid activation function that normalizes the feature map of each channel;

[0029] e c Represents the weight of the output of the incentive calculation.

[0030] According to one embodiment of the present invention, when the bottleneck module (neck) of the model is lightweight, a lightweight convolution GSConv module is used instead of ordinary convolution, which can effectively reduce the amount of model calculation and achieve information exchange between feature channels with lower time loss.

[0031] According to one embodiment of the present invention, a GSBottleneck module and a single aggregated VoV-GSCSPC fusion module are used based on the GSConv module.

[0032] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 This is the structural principle diagram of the YOLOv8-BL network model;

[0036] Figure 2 This is the structural principle diagram of the SE channel attention mechanism module;

[0037] Figure 3 This is the structural schematic diagram of the GSConv module;

[0038] Figure 4 This is the structural schematic diagram of the GSBottleneck module;

[0039] Figure 5 It is the structural schematic diagram of a single aggregated VoV-GSCSPC fusion module;

[0040] Figure 6 It is a performance comparison diagram of a perception algorithm of an intelligent vehicle light self-closed loop and assisted driving system of the present invention and other benchmark model algorithms;

[0041] Figure 7 It is a visualization diagram of the reasoning results of the perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The following describes in detail a perception algorithm of an intelligent vehicle light self-closing loop and assisted driving system according to an embodiment of the present invention with reference to the accompanying drawings.

[0044] See Figures 1 to 7 The perception algorithm of the intelligent headlight self-closed loop and assisted driving system of the present invention is mainly based on high-mounted brake light state detection, that is, detecting the state of the high-mounted brake lights of the vehicle in front and inputting it into the decision-making and control system of the vehicle and the intelligent headlight; using the YOLOv8-BL network model re-optimized and rebuilt based on the YOLOv8n network model to complete the vehicle high-mounted brake light detection in the front view image collected by the front view camera in the intelligent assisted driving and intelligent headlight self-closed loop system.

[0045] The YOLOv8n network model structure is mainly divided into four parts: input for data enhancement, backbone network (BackBone) for image feature extraction, bottleneck module (Neck) for feature fusion, and decoupling head for separating classification from detection head. The structure of the YOLOv8-BL model described in the present invention is as follows: Figure 1 shown.

[0046] Adding the SE channel attention mechanism module to the YOLOv8n network model structure makes the YOLOv8n network model more suitable for feature extraction of key areas of the image. The SE channel attention mechanism module adaptively adjusts the importance of each channel in the feature map by learning the attention of different channels in the image.

[0047] The lightweight convolution GSConv module and VoV-GSCSPC module are used to replace the CrossStage Partial Network Fusion (C2f) fusion module in the YOLOv8n network model. The local and global structural features are captured by aggregating multi-layer neighborhood information, and the input feature maps are grouped for processing. Each group performs convolution operations and parallel calculations independently.

[0048] The channel attention SE channel attention mechanism module is used to extract image features. The SE channel attention mechanism module enhances the extraction of image difference information by weighted processing of features of different channels of the image, enhances the learning of the key feature areas of the target by the YOLOv8-BL network model, and helps to improve the detection accuracy of the YOLOv8-BL network model for the high-mounted brake light area in the image.

[0049] The YOLOv8 network model is the latest model in the YOLO series. It designs the network model architecture based on the YOLOv5 network model, introduces a new structure, and further improves performance and scalability. Based on depth and width, it can be divided into five networks: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. The number of model parameters and the amount of calculation are greatly increased with the improvement of accuracy to meet the needs of different scenarios.

[0050] The YOLOv8n network model uses the Spatial Pyr-amid Pooling Fast (SPPF) feature image pyramid pooling module in the backbone network, which reduces the amount of computation and increases the receptive field to a certain extent, but due to non-multi-scale prediction, it is relatively weak in detecting the state of high-mounted brake lights. High-mounted brake lights usually have less pixel information in images, which makes it difficult to accurately locate and classify. The SE channel attention mechanism module can adaptively adjust the importance of each channel in the feature map by learning the attention of different channels in the image. By adding the SE channel attention mechanism module, the model is more suitable for feature extraction of key areas of the image.

[0051] The bottleneck module of the YOLOv8n network model adopts the Efficient Layer Aggregation Networks (ELAN) structure in YOLOv7, and adopts the Cross Stage Partial Network Fusion (C2f) fusion module, through more branch cross-layer connections, to form a neural network model with stronger feature representation capabilities, while enriching the gradient flow of the neural network model; however, the neural network model structure is relatively large and deep, and requires a longer training time to reach convergence. Compared with other target detection algorithms, the training speed is slower. Therefore, the present invention uses a lightweight convolution GSConv module and a VoV-GSCSPC module to capture local and global structural features by aggregating multi-layer neighborhood information, and groups the input feature map for processing. Each group independently performs convolution operations and parallel calculations. The above calculation method reduces the number of parameters and the calculation burden, and improves the calculation efficiency of the model.

[0052] Since the high-mounted brake lights are densely distributed and small in the front-view image dataset of the intelligent assisted driving and intelligent headlight self-closed loop system, when the basic YOLOv8n model is used for target detection and classification, the model has a weak feature extraction capability for the high-mounted brake lights target and low detection accuracy, which cannot meet the detection accuracy requirements of the intelligent assisted driving and intelligent headlight self-closed loop system. To solve the above problem, the present invention embeds the SE channel attention mechanism module into the backbone network (BackBone) to form a new feature extraction network SE-BackBone. The SE channel attention mechanism module is as follows: Figure 2 As shown in the figure. In the SE-BackBone network, the output feature map Q1 of the SPPF layer is passed to the SE channel attention mechanism module as the input feature, and is compressed by the global average pooling (GAP) layer and stimulated by two fully connected layers (FC) to obtain the channel weighted feature map Q2. The feature map Q2 is multiplied by the output feature map Q1 of the SPPF layer channel by channel to recalibrate the importance of the input feature, and the output feature map Q3 of the SE channel attention mechanism module is obtained.

[0053] The SE channel attention mechanism module only focuses on information that is more relevant to the current task, and suppresses useless information by adaptively learning the relevance and importance of feature channels to improve the representation ability of the network model. In the YOLOv8-BL network model, the channel attention SE channel attention mechanism module is used to extract image features. The SE channel attention mechanism module enhances the extraction of image difference information and the model's learning of the target's key feature areas by weighted processing of features of different channels of the image, which helps to improve the model's detection accuracy of the high-mounted brake light area in the image.

[0054] The mapping relationship of the SE channel attention mechanism module is shown in formula (1), formula (2) and formula (3):

[0055]

[0056] L we (u c , e c )=e c u c (3)

[0057] The symbols in formula (1), formula (2) and formula (3) mean:

[0058] L sq Indicates compressed calculation;

[0059] L ex represents incentive calculation;

[0060] L we represents weighted calculation;

[0061] H represents the height of the input feature Q1;

[0062] W represents the width of the input feature Q1;

[0063] i and j represent the continuous values ​​of pixels. The value range of i is [0, H], and the value range of j is [0, W];

[0064] u c Represents each feature channel;

[0065] s represents the number of channels after compression calculation;

[0066] Represents the nonlinear activation function Relu;

[0067] W0 represents the total parameters of the fully connected (FC) calculation;

[0068] W1, W2, and g represent the parameters of the fully connected calculation of each layer;

[0069] θ represents the Sigmoid activation function that normalizes the feature map of each channel;

[0070] e c Represents the weight of the output of the incentive calculation.

[0071] When the bottleneck module (neck) of the model is lightweight, the present invention adopts a lightweight convolution GSConv module ( Figure 4 ) replaces the ordinary convolution to achieve information exchange between feature channels with lower time loss. Specifically, in the bottleneck module (Neck) of the yolov8-BL network of the present invention, GSConv convolution is used to replace the ordinary Conv structure in the original yolov8n, which can effectively reduce the amount of model calculation. The network structure of GSConv is different from that of ordinary Conv, which can be referred to Figure 3 .

[0072] The complexity of convolution calculation is usually measured in GFLOPs. The time complexity calculation formula of the GSConv module is:

[0073]

[0074] The symbols in formula (4) mean:

[0075] w represents the width of the output feature map;

[0076] h represents the height of the output feature map;

[0077] K1 and K2 represent the size of the convolution kernel;

[0078] C1 represents the number of channels of each convolution kernel, which is also equal to the number of channels of the input feature map;

[0079] C2 represents the number of channels of the output feature map;

[0080] O stands for Big O method (the time complexity of an algorithm is usually measured using the Big O method).

[0081] The advantages of the GSConv module are evident in lightweight detectors, which use aggregated multi-layer neighborhood information to capture local and global structural features, and use uniform mixing operations to add information generated by standard convolutions to information generated by depthwise separable convolutions. If the GSConv module is used in all stages of the YOLOv8-BL network model, the number of network layers will be deeper, increasing the resistance to data flow and significantly increasing the inference time. Therefore, the GSConv module is only used in the bottleneck module of the YOLOv8-BL network model. At this stage, it is most appropriate to use the GSConv module to process the connected feature maps, with less redundant and repeated information, and no need to compress the information.

[0082] In the YOLOv8-BL network model, the GSBottleneck module and the single aggregation VoV-GSCSPC fusion module are further used on the basis of the GSConv module, such as Figure 5 and Figure 6 As shown in the figure, the GSConv module replaces the ordinary convolution operation of the original bottleneck module, and the single aggregated VoV-GSCSPC fusion module replaces the C2f module of the original bottleneck module, and the GSBottleneck module replaces the Bottleneck module. The design of the lightweight module reduces the computational complexity of the YOLOv8-BL network model while maintaining the detection accuracy of the YOLOv8-BL network model. It should be supplemented that the core components of the bottleneck module of the YOLOv8-BL of the present invention are GSConv, GSBottleneck, and VoV-GSCSPC, and these three components have the following relationship: GSBottleneck contains GSConv, and VoV-GSCSPC contains GSBottleneck. The core components of the original YOLOv8n bottleneck module are Conv, Bottleneck, and C2f, and these three components have the following relationship: C2f contains Bottleneck, and Bottleneck contains Conv.

[0083] See Figure 6, the unit of calculation amount is billion floating point operations, and the unit of parameter amount is M (M represents one million). Among the four models of YOLOv8l, YOLOv8m, YOLOv8s and YOLOv8n, the mAP@0.5 evaluation index (average precision) is the highest in YOLOv8n, the least calculation amount is YOLOv8n, and the least parameter amount is also YOLOv8n. By comparison, the YOLOv8-BL network model proposed in the present invention reduces the amount of calculation by 12.64% and the amount of parameters by 12.5% ​​compared with the YOLOv8n benchmark model, but the mAP@0.5 evaluation index (average precision) of the YOLOv8-BL network model is improved by 3.9% compared with the YOLOv8n benchmark model. The improved model has greatly reduced the amount of parameters and calculation, and has stronger performance, which is more suitable for the deployment and application of the high-mounted brake light state detection model on small devices.

[0084] The perception algorithm of the intelligent self-closed loop and assisted driving system of the present invention can enable the intelligent assisted driving and intelligent self-closed loop system of the intelligent headlight to obtain richer environmental information, more accurately predict the future state, make humanized intelligent headlight and vehicle control decisions, reduce sensory conflicts, and improve riding comfort and driving safety. By deeply innovating and optimizing the existing YOLOv8 algorithm, the specific beneficial effects are as follows: (1) Adding the SE channel attention mechanism module, the model adaptively extracts features from the key areas of the vehicle to improve detection accuracy; (2) Introducing lightweight convolution in the Conv module, replacing ordinary convolution with the GSConv module, effectively reducing the amount of model calculation; (3) Based on the GSConv module, a single aggregation module VoV-GSCSPC is designed to optimize the model structure. This method maintains information exchange between channels with lower time loss, which improves detection accuracy while reducing the amount of model calculation.

[0085] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A perception algorithm for an intelligent vehicle light self-closed loop and assisted driving system, characterized in that: It is mainly based on high-mounted brake light status detection, that is, detecting the status of the high-mounted brake lights of the vehicle in front and inputting it into the decision-making and control system of the vehicle and intelligent headlights; using the YOLOv8-BL network model re-optimized and rebuilt based on the YOLOv8n network model to complete the vehicle high-mounted brake light detection in the front view image collected by the front camera in the intelligent assisted driving and intelligent headlight self-closed loop system.

2. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 1, characterized in that: The YOLOv8n network model structure includes an input for data enhancement, a backbone network (BackBone) for image feature extraction, a bottleneck module (Neck) for feature fusion, and a decoupling head for separating classification from detection head.

3. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 2 is characterized by: Adding the SE channel attention mechanism module to the YOLOv8n network model structure makes the YOLOv8n network model more suitable for feature extraction of key areas of the image. The SE channel attention mechanism module adaptively adjusts the importance of each channel in the feature map by learning the attention of different channels in the image.

4. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 3 is characterized by: The lightweight convolution GSConv module and VoV-GSCSPC module are used to replace the Cross Stage Partial Network Fusion (C2f) fusion module in the YOLOv8n network model. The local and global structural features are captured by aggregating multi-layer neighborhood information, and the input feature maps are grouped for processing. Each group performs convolution operations and parallel calculations independently.

5. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 4 is characterized by: The YOLOv8n network model uses the Spatial Pyr-amid Pooling Fast (SPPF) feature image pyramid pooling module in the backbone network.

6. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 5, characterized in that: Embed the SE channel attention mechanism module into the backbone network (BackBone) A new feature extraction network SE-BackBone is formed. In the SE-BackBone network, the output feature map Q1 of the SPPF layer is passed to the SE channel attention mechanism module as the input feature, and is compressed by the global average pooling layer and excited by two fully connected layers to obtain the channel weighted feature map Q2. The feature map Q2 is multiplied by the output feature map Q1 of the SPPF layer channel by channel to recalibrate the importance of the input feature, and the output feature map Q3 of the SE channel attention mechanism module is obtained.

7. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 6 is characterized by: The channel attention SE channel attention mechanism module is used to extract image features. The SE channel attention mechanism module enhances the extraction of image difference information by weighted processing of features of different channels of the image, enhances the learning of the key feature areas of the target by the YOLOv8-BL network model, and helps to improve the detection accuracy of the YOLOv8-BL network model for the high-mounted brake light area in the image.

8. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 7, characterized in that: The mapping relationship of the SE channel attention mechanism module is shown in formula (1), formula (2) and formula (3): L we (the c ,and c )=and c the c (3) The symbols in formula (1), formula (2) and formula (3) mean: L sq Indicates compressed calculation; L ex represents incentive calculation; L we represents weighted calculation; H represents the height of the input feature Q1; W represents the width of the input feature Q1; i and j represent the continuous values ​​of pixels. The value range of i is [0, H], and the value range of j is [0, W]; u c Represents each feature channel; s represents the number of channels after compression calculation; Represents the nonlinear activation function Relu; W0 represents the total parameters of the fully connected (FC) calculation; W1, W2, and g represent the parameters of the fully connected calculation of each layer; θ represents the Sigmoid activation function that normalizes the feature map of each channel; e c Represents the weight of the output of the incentive calculation.

9. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 8, characterized in that: When lightweighting the bottleneck module (neck) of the model, using the lightweight convolution GSConv module instead of the ordinary convolution can effectively reduce the amount of model calculation and achieve information exchange between feature channels with lower time loss.

10. The perception algorithm of the intelligent vehicle light self-closed loop and assisted driving system according to claim 9, characterized in that: Based on the GSConv module, the GSBottleneck module and the single aggregation VoV-GSCSPC fusion module are used.

Images