Braking control method and system for nighttime vehicle target detection based on improved KSC-YOLOV5

Abstract

The present invention provides a braking control method and system for nighttime vehicle target detection based on improved KSC‑YOLOV5, including: constructing a driving simulation space according to the current driving environment of the vehicle at night; inputting the front image of the vehicle photographed in the current driving environment into the improved KSC‑YOLOV5 model for target detection; controlling the action of the distance measurement device to obtain distance measurement data according to the target detection result; constructing an interference target model in the driving simulation space according to the target detection result and the distance measurement data; analyzing the behavior of the interference target model, and issuing a braking reminder and / or controlling the vehicle braking according to the analysis result. The braking control method and system for nighttime vehicle target detection based on improved KSC‑YOLOV5 of the present invention improve the accuracy and effectiveness of braking control of vehicles at night, and help improve the safety of nighttime driving.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle braking control, and in particular to a braking control method and system for nighttime vehicle target detection based on improved KSC-YOLOV5. Background Art

[0002] With the annual growth of car ownership, traffic safety issues are becoming increasingly serious. In nighttime environments, due to low light levels, visibility is poor and drivers are prone to misjudging the environment. Currently, there are few applications of automotive automatic driving systems and advanced driving assistance systems (ADAS) in nighttime conditions. Improving the accuracy and effectiveness of braking control of vehicles at night will help improve the safety of nighttime driving. Summary of the invention

[0003] One of the purposes of the present invention is to provide a braking control method and system for nighttime vehicle target detection based on improved KSC-YOLOV5, so as to improve the accuracy and effectiveness of braking control of nighttime vehicles and contribute to improving the safety of nighttime driving.

[0004] An embodiment of the present invention provides a braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5, comprising:

[0005] Construct a driving simulation space based on the current driving environment of the vehicle at night;

[0006] Input the front image of the vehicle taken in the current driving environment into the improved KSC-YOLOV5 model for target detection;

[0007] According to the target detection result, the distance measurement device is controlled to obtain distance measurement data;

[0008] Based on the target detection results and distance measurement data, an interference target model is constructed in the driving simulation space;

[0009] The behavior of the interference target model is analyzed, and a braking reminder is issued and / or the vehicle braking is controlled based on the analysis result.

[0010] Preferably, the steps for constructing the driving simulation space are as follows:

[0011] Retrieving the road parameters of the road where the vehicle is located from the road information storage base according to the positioning information of the vehicle;

[0012] According to the road parameters, the driving simulation space parameters and the initial model space template are retrieved from the simulation space determination library;

[0013] The driving simulation space is constructed according to the retrieved driving simulation space parameters and the initial model space template.

[0014] Preferably, the improved KSC-YOLOV5 model includes: a SPFF module.

[0015] Preferably, the Neck module of the improved KSC-YOLOV5 model includes: a path aggregation network and a feature pyramid network.

[0016] Preferably, the improved KSC-YOLOV5 model adopts the S IOU loss function, and the formula of the loss function is as follows:

[0017]

[0018] Where, SIOU loss represents the SIOU loss function; IOU represents the IOU loss; Δ is the distance loss under the condition of considering the angle loss; Ω represents the shape loss; B and B GT They represent the area of ​​the true bounding box and the area of ​​the predicted bounding box respectively.

[0019] Preferably, the control steps of the distance measuring device are as follows:

[0020] Using a preset feature extraction template, extract features from the target detection results and generate an analysis data set based on the extracted features;

[0021] The control parameters are determined based on the analysis data set and a pre-configured control parameter determination library.

[0022] Preferably, the behavior analysis step includes:

[0023] According to the position change of the interference target model at the previous moment and the current moment, the moving direction of the interference target model is analyzed;

[0024] Based on the interference target model and moving direction at the current moment, the trajectory of the interference target model is predicted;

[0025] Determine whether the predicted trajectory intersects with the vehicle's trajectory, and output a braking reminder if an intersection occurs;

[0026] Based on the preset trajectory, it is determined whether the interference target model will collide with the vehicle model corresponding to the vehicle, and when so, the vehicle is controlled to brake.

[0027] The present invention also provides a braking control system for nighttime vehicle target detection based on improved KSC-YOLOV5, comprising: a space construction unit, a target detection unit, a distance measurement unit, a target construction unit and a behavior analysis unit;

[0028] Among them, the space construction unit constructs a driving simulation space according to the current driving environment of the vehicle at night; the target detection unit inputs the front image of the vehicle taken in the current driving environment into the improved KSC-YOLOV5 model for target detection; the distance measurement unit controls the action of the distance measurement device according to the target detection result to obtain the distance measurement data; the target construction unit constructs the interference target model in the driving simulation space according to the target detection result and the distance measurement data; the behavior analysis unit analyzes the behavior of the interference target model, and issues a braking reminder and / or controls the vehicle braking according to the analysis result.

[0029] Preferably, the space construction unit performs the following operations:

[0030] Retrieving the road parameters of the road where the vehicle is located from the road information storage base according to the positioning information of the vehicle;

[0031] According to the road parameters, the driving simulation space parameters and the initial model space template are retrieved from the simulation space determination library;

[0032] The driving simulation space is constructed according to the retrieved driving simulation space parameters and the initial model space template.

[0033] Preferably, the control steps of the distance measuring device are as follows:

[0034] Using a preset feature extraction template, extract features from the target detection results and generate an analysis data set based on the extracted features;

[0035] The control parameters are determined based on the analysis data set and a pre-configured control parameter determination library.

[0036] 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 can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0039] Figure 1 A schematic diagram of a braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5 in an embodiment of the present invention;

[0040] Figure 2This is a CSP structure diagram in an embodiment of the present invention;

[0041] Figure 3 It is a structural diagram of SPP and SPPF in an embodiment of the present invention;

[0042] Figure 4 This is a structural diagram of KSC-YOLOv5 in an embodiment of the present invention;

[0043] Figure 5 Schematic diagram of the relationship between the SIOU predicted box and the real box in an embodiment of the present invention;

[0044] Figure 6 This is a structural diagram of a CARAFE module in an embodiment of the present invention;

[0045] Figure 7 Schematic diagram of a selective convolution kernel in an embodiment of the present invention;

[0046] Figure 8 Schematic diagram of a braking control system for nighttime vehicle target detection based on improved KSC-YOLOV5 in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0048] The embodiment of the present invention provides a braking control method based on night vehicle target detection of improved KSC-YOLOV5, such as Figure 1 As shown, including:

[0049] Step 1: construct a driving simulation space based on the current driving environment of the vehicle at night;

[0050] Step 2: Input the front image of the vehicle taken in the current driving environment into the improved KSC-YOLOV5 model for target detection;

[0051] Step 3: According to the target detection result, the distance measurement device is controlled to obtain distance measurement data;

[0052] Step 4: Construct an interference target model in the driving simulation space based on the target detection results and distance measurement data;

[0053] Step 5: Analyze the behavior of the interference target model, and issue a braking reminder and / or control vehicle braking based on the analysis results.

[0054] The steps for constructing the driving simulation space are as follows:

[0055] According to the positioning information of the vehicle, the road parameters of the road where the vehicle is located are retrieved from the road information storage library; the road information storage library is pre-configured, and the road parameters of the roads at various locations are stored therein;

[0056] According to the road parameters, the driving simulation space parameters and the initial model space template are retrieved from the simulation space determination library; the simulation space determination library is also configured in advance; wherein the driving simulation space parameters and the initial model space template are correspondingly associated with the road parameters;

[0057] The driving simulation space is constructed based on the retrieved driving simulation space parameters and the initial model space template. The driving simulation space parameters include: the size of the simulation space; the initial model space template includes: the templates of various fixed objects in the space (roadside trees, signboards, curbs, etc.).

[0058] Among them, the improved KSC-YOLOV5 model includes: SPFF module.

[0059] Among them, the Neck module of the improved KSC-YOLOV5 model includes: path aggregation network and feature pyramid network.

[0060] Among them, the improved KSC-YOLOV5 model adopts the S IOU loss function, and the formula of the loss function is as follows:

[0061]

[0062] Where, SIOU loss represents the SIOU loss function; IOU represents the IOU loss; Δ is the distance loss under the condition of considering the angle loss; Ω represents the shape loss; B and B GT They represent the area of ​​the true bounding box and the area of ​​the predicted bounding box respectively.

[0063] The working principle and beneficial effects of the above technical solution are:

[0064] The framework of the YOLO series target detection can usually be divided into the following parts: input, backbone network, Neck network and output.

[0065] The backbone of YOLOv5 contains several important modules, namely the bottleneck cross-stage partial network (CSP) and spatial pyramid pooling (SPPF).

[0066] The input of YOLOv5 mainly includes three key parts: Mosaic data enhancement, adaptive anchor box calculation and adaptive image scaling. The principle of the Mosaic data enhancement method is similar to the CutMix method. But the difference is that CutMix uses two pictures for operation, while Mosaic data enhancement stitches four different images into one image by random cropping, scaling and arbitrary arrangement as the input of the detector. This method significantly enhances the overall accuracy and robustness of the algorithm, thereby improving the model's ability to detect objects in complex and diverse environments. Adaptive anchor box calculation means that during the model training process, the optimal anchor box size can be automatically calculated without the need to calculate the anchor box size separately on different data sets. In this way, the model can better adapt to the characteristics of different data sets and improve the generalization ability of the model. In general target detection algorithms, the input end uniformly scales different pictures to one size, which may introduce some redundant black edges, thereby affecting the inference speed. The method of adaptive image scaling is to fill the least pixels as much as possible and scale the original image to a standard size to reduce redundant information, which is conducive to improving the inference speed of the network. These optimization measures make YOLOv5's input processing more flexible and efficient, and the model can detect target objects more accurately in complex and diverse environments.

[0067] CSP module: Due to the problem of repeated gradient calculation during network optimization, the inference process has too much calculation. To solve this problem, YOLOv5 borrowed the idea of ​​CSPNet and applied the CSP module in multi-scale feature extraction. This module is conducive to enhancing the learning ability of the network and can reduce the amount of calculation while ensuring accuracy. YOLOv5 designed two CSP modules, the structure is as follows Figure 2 As shown in the figure, CSP1_X is used for the backbone network, while CSP2_X is used for the Neck network. CSP1_X first splits the feature map into two parts according to the channel, one part performs conventional convolution operations, and the other part uses the idea of ​​residual network to construct residual components, and finally merges the two parts to obtain a new feature map. This design effectively avoids repeated calculations of gradient values ​​and improves the model reasoning speed. At the same time, the CSP structure has a residual component, which can enhance the gradient value during the back propagation process, especially when the backbone network is deep, alleviating the gradient disappearance problem and enhancing the network's feature extraction ability. CSP2_X uses a convolutional layer instead of a residual component to divide the input feature map into two parts, calculate them separately and then fuse them, which can retain more image information.

[0068] Spatial pyramid pooling SPPF: The structures of SPP module and SPPF module are as follows Figure 2As shown in the figure, in the SPP module, the feature maps are input in parallel into the maximum pooling layers with convolution kernel sizes of 5×5, 9×9, and 13×13, and then the four feature maps with different perceptions are spliced ​​together. In contrast, the SPPF module uses a maximum pooling layer with a serial input convolution kernel size of 5×5, and then performs feature fusion. It is worth noting that since two 5×5 maximum pooling layers have the same effect as one 9×9 maximum pooling layer, and the calculation results of three 5×5 maximum pooling layers and one 13×13 maximum pooling layer are also the same. Therefore, the use of the SPPF structure can achieve the same effect, while the model has a smaller amount of calculation and higher efficiency.

[0069] NECK module: As feature extraction proceeds, some local information of the image may be lost. In order to make full use of the feature maps of different network layers, the Neck network is introduced to fuse these features so that the processed features can be better input into the Head layer for classification and positioning. Since the feature map extracted by Backbone is large in length and width, and the final classification and positioning output result dimension is small, the feature dimension needs to be compressed, so this part of the structure is called Neck. The Neck structure helps to more effectively utilize the features extracted by Backbone, further fuse and process them, and then transmit them to the output end for classification and positioning.

[0070] The Neck module of YOLOv5 adopts a composite structure of path aggregation network (PAN) and feature pyramid network (FPN). FPN improves the detection effect of small targets by integrating high- and low-level features, while PANet transfers features in a bottom-up manner based on FPN to achieve the fusion of low-level features and high-level features. This design enables the top-level feature map to effectively utilize the location information of the underlying features, thereby avoiding the problem that the underlying features cannot affect the high-level features caused by the single top-down feature transfer of FPN. This structure organically combines these two methods, aggregates parameters of different detection layers, and further improves the detection effect of dense targets and multi-scale targets.

[0071] HEAD is the final detection part of Yolo v5. The main task of the output end is to compress and fuse the features extracted from the backbone through the neck part, and then classify and predict the tasks. In order to effectively evaluate the position loss of the target box and the prediction box, the model adopts the CIOU loss function to avoid the problem of slow convergence caused by the non-overlapping bounding boxes. The YOLOv5 model has three scale feature maps, namely 80x80, 40x40, and 20x20 grids. Different grids correspond to different sizes of detection targets. By using the NMS algorithm, the model can effectively eliminate redundant prediction boxes that do not meet the threshold conditions, thereby enhancing the target recognition ability and completing the entire target detection process.

[0072] Method: In order to solve the problem of low accuracy of the network model, the Neck of YOLOv5 was redesigned. Using CARAFE (position as Figure 3 The traditional upsampling operator is optimized. CARAFE is lightweight, fast, introduces little computational overhead, and is easy to integrate into modern network architectures. In order to improve the convolution capability, the Backbone part of YOLOv5 is improved. SKNet is added before and after SPPF to improve the feature extraction capability of the target and obtain more vehicle feature information. At the same time, SIOU is introduced to redefine the regression loss and promote model convergence.

[0073] SIOU: Taking into account the aspect ratio of the bounding box, the overlapping area between the bounding boxes, and the center point distance between the bounding boxes, the original YOLOv5 target detection algorithm uses CIOU as the position regression loss function, which is calculated as follows:

[0074]

[0075] Among them, b, bgt represent the distance between the center point of the model predicted target box and the actual real box; p represents the Euclidean distance between the center points of the predicted box and the real box; c represents the diagonal distance of the minimum closure area that can simultaneously contain the model predicted target box and the actual real box; a represents the parameter of the balance ratio; v represents the consistency of the aspect ratio of the model predicted target box and the actual real box. Among them, the formulas of a and v are:

[0076]

[0077] CI oU optimizes the distance, overlap, scale and aspect ratio between the predicted box and the true box in the target detection algorithm, thereby improving accuracy and stability. However, CI oU does not consider the direction mismatch between the true box and the predicted box, which may cause the model to be unstable during training, affect the convergence and detection efficiency of the model, and make the predicted box "wander around" during the training process. To solve this problem, this paper uses SIOU as the loss function, introduces the vector angle between the true box and the predicted box, and considers the vector angle between the expected regression. By redefining the penalty metric, the shape information of the target box is more effectively utilized. This improvement helps to promote the fitting of the true box and the predicted box, so that the model can locate the defect position more accurately, and accelerates the convergence process of the model, thereby improving the model training speed and the accuracy of reasoning. The SIOU loss function consists of four cost functions: angle loss, distance loss, shape loss, and IoU loss (Anglecost, Distancecost, Shapecost, and IoUcost). The formula of SIOU loss is as follows:

[0078]

[0079] Where: IOU is the IOU loss, Λ is the angle loss, Δ is the distance loss under the condition of considering the angle loss, Ω is the shape loss, B and B GT They represent the area of ​​the true bounding box and the area of ​​the predicted bounding box respectively.

[0080] Using S IOU loss and adding angle cost calculation effectively reduces the total degrees of freedom of the loss, redefines the regression distance loss, and effectively helps model training, which not only promotes the convergence of the model, but also effectively improves the accuracy of the network model.

[0081] CARAFE: In the object detection network, upsampling operation is usually used to expand feature maps to achieve matching between feature maps at different levels. In the baseline network YOLOv5s, the nearest neighbor interpolation method is used for upsampling operation, but the receptive field is small, only 1×1, and it is very easy to show jagged edges or lines at the edges or lines of the feature map after upsampling, which has a certain impact on the positioning and recognition of defect targets, and thus cannot achieve the optimal detection effect. Therefore, a lightweight and efficient CARAFE operator is introduced to upsample the feature map. CARAFE uses the underlying content information to predict the reorganization kernel and reassemble the predefined nearby area features. This method can apply adaptive and optimized reorganization kernels at different locations, significantly improving the quality and efficiency of upsampling. The CARAFE module consists of two parts: the kernel prediction module and the feature perception reorganization module. In the kernel prediction module, assuming the upsampling ratio is σ and the input feature map χ is of size H×W×C, first, the number of channels is compressed from C to Cm to reduce the computational complexity of subsequent operations and improve the upsampling efficiency; then the feature content is encoded and upsampling prediction is performed. Since different position information uses different upsampling kernels, let the upsampling kernel be kup×kup, the number of input channels be Cm, and the output

[0082] The number of channels is σ2k2up, and then the feature image is expanded in the channel dimension to obtain an upsampling kernel of size σH×σW×k2up; finally, the predicted upsampling kernel is normalized by the Softmax operation to ensure that the weight sum of the convolution kernel is 1. The feature-aware reorganization module mainly involves mapping and dot product operations. First, each position of the output feature map is mapped and back-mapped to the input feature map to form a new feature map N(χ,kup). Then, for the center of each position,

[0083] Then, the kup×kup region at the center of each position is taken out for dot product operation with the reorganized kernel Wr, so that the same upsampling kernel can share the upsampling kernel information of different channels at the same position, and finally the upsampling feature map χ' of size σH×σW×C is obtained. This process realizes the recombination of feature maps through dot product operation, so that the upsampling kernel can be shared between different channels, and a more adaptable upsampling feature map is obtained, which provides an effective means to further improve the performance of the model. Through the collaborative work of these two modules, the CARAFE module can use adaptive and optimized reorganized kernels at different positions to improve the quality and efficiency of upsampling. This design helps to overcome the problems in traditional upsampling methods, especially in target detection tasks, and has significant improvements in accurately locating and identifying defective targets.

[0084] SKNet: In night scenes, vehicles are often affected by occlusions of different shapes and sizes. Therefore, how to obtain dynamic weight adjustment information so that the detector can adapt to various complex environments is a problem worthy of in-depth study. Therefore, the SKNet (Selective Kernel) attention mechanism is adopted in KSCYo lo. SKNet (Selective Kernel Network) is an image classification model based on a deep residual network structure. Its core idea is to solve the information overlap and redundancy problems in the model by introducing an adaptive attention mechanism, thereby further improving the performance of the model. The structural diagram of the model is shown in the figure. Figure 6 As shown in Figure 1, it mainly consists of three parts, namely "Split", "Fuse" and "Select" operations, as follows. In the Split part, the purpose is to generate multiple paths with different convolution kernel sizes, corresponding to different receptive sizes of neurons. For a given feature map X∈RH′×W′×C′, convolution kernels of sizes 3×3 and 5×5 are used for convolution operations. Batch normalization is performed in the two branches, and ReLU is used as the activation function. The feature maps obtained after transformation are represented as U^ and

[0085] The purpose of the fuse part is to aggregate information from different paths so that these paths carry information of different scales and pass it to the neurons in the next layer. First, the results from multiple branches are fused by summing the elements, as shown in the following formula:

[0086]

[0087] Then, global information is generated by global average pooling to generate channel-wise statistics.

[0088] ∈RC, the cth element of s is obtained by shrinking U to the spatial dimension H×W:

[0089]

[0090] In addition, in order to reduce the dimension to improve efficiency, a compact feature z∈Rd×1 is created. As shown in the following formula, δ is the ReLU function, B represents batch normalization, and W∈Rd×C. d represents the impact on model efficiency, and the reduction ratio r is used to control its value. L represents the minimum value of d.

[0091] z=F fc (s) = δ(B(W s ))

[0092] d=max(C / r,L)

[0093] The Select part multiplies the two matrices a and b with the feature map obtained in the Split part and then performs a sum operation to obtain the feature map V. The generation method of attention weights ac and bc is as follows:

[0094]

[0095] Among them, A, B∈RC×d, a, B represent the Soft attention vector. Ac∈R1×d represents the cth row of A, ac is the cth element of a, and the same is true for BC and bC. In the case of two branches, the matrix B is redundant because ac+bc=1. The final feature map V is obtained by the attention weights on various kernels:

[0096] V c =a c ·U c +b c ·U c ,a c +b c =1

[0097] Among them, V = [V1, V2, ..., VC], Vc ∈ RH × W. This design of SKNet enables the network to dynamically adapt to different receptive field sizes, so as to better capture the local and global information in the image and improve the performance and generalization ability of the model.

[0098] In order to achieve accurate measurement of the target position after target recognition, in one embodiment, the control steps of the distance measurement device are as follows:

[0099] Using a preset feature extraction template, feature extraction is performed on the target detection result to generate an analysis data set based on the extracted features; the extracted features include: parameters representing the vehicle type of the target, parameters representing the coordinates of the boundary of the target in the image, etc.;

[0100] The control parameters are determined based on the analysis data set and the pre-configured control parameter determination library. The control parameter determination library is configured by professionals in advance, and the analysis data set and the control parameters are correspondingly associated in the library; the control parameters determined by the control parameter determination library mainly include: the measurement angle adjustment parameters of the measuring device, etc.

[0101] In order to achieve intelligent vehicle braking, in one embodiment, the behavior analysis step includes:

[0102] According to the position change of the interference target model at the previous moment and the current moment, the moving direction of the interference target model is analyzed;

[0103] Based on the interference target model and moving direction at the current moment, the trajectory of the interference target model is predicted; according to the moving direction and moving speed, multiple prediction points are determined, the direction between adjacent prediction points corresponds to the moving direction, and the distance between them is the moving speed multiplied by the time difference corresponding to the prediction point; each prediction point is connected to form a predicted trajectory;

[0104] Determine whether the predicted trajectory intersects with the vehicle's trajectory. If so, output a braking reminder. The vehicle trajectory is constructed in the same way as the predicted trajectory. The intersection of the two trajectories refers to whether there is an intersection between the two trajectories.

[0105] Based on the preset trajectory, determine whether the interference target model will collide with the vehicle model corresponding to the vehicle. If so, control the vehicle to brake. Associate the points on the two trajectories at the same time, and then determine the model area of ​​each point based on the associated points. When the model areas overlap, it is determined that a collision will occur.

[0106] In one embodiment, the behavior analysis step further includes:

[0107] According to the position change of the interference target model at the previous moment and the current moment, the moving direction and moving speed of the interference target model are analyzed as the moving parameters at the previous moment;

[0108] According to the positioning of the vehicle and the relative position between the interference target model and the vehicle at each moment, the position of the interference target model at each moment in the pre-configured road condition data distribution map is determined and the corresponding road condition data is retrieved; the road condition data includes: road surface type, road surface flatness, etc.;

[0109] Extract the road condition area image corresponding to the target corresponding to the interference target model in the front image of the vehicle at each time, input the pre-configured road condition analysis model, and obtain real-time road condition data; the real-time road condition data includes: whether there are obstacles or pits, obstacle or pit size parameters, etc.;

[0110] According to the movement parameters at each moment in the previous N moments, the corresponding retrieved road condition data and the real-time road condition data, an analysis data set for target stability analysis is constructed; each data is quantified and arranged in sequence to form an analysis data set, for example: the first row of data and the second row of data in the constructed analysis data set are the quantified data of the movement parameters at each moment; the third row and the fourth row are the quantified data of the road surface type and the road surface flatness; the fifth row to the eighth row are the quantified data of the real-time road condition data, etc.;

[0111] Matching the analysis data set with the standard data set associated with each behavior analysis result in the pre-configured behavior analysis library; the behavior analysis library is constructed by professionals in advance, and the behavior analysis results are associated with the standard data set in the library one by one. The standard data set and the analysis data set can be matched by calculating the similarity between the two and then comparing the calculated similarity with a preset threshold. When it is greater than or equal to the threshold and is the maximum, it is determined to be a match;

[0112] Extract behavioral analysis results associated with a standard dataset that matches the analysis dataset.

[0113] The behavior of other vehicles in the image in front of the vehicle is determined from the behavior analysis results. When the behavior analysis results show that the vehicle is out of control or has poor stability, braking control is performed.

[0114] The rule for extracting the road condition area image is to map the moving direction to the image in front of the vehicle to determine the center offset vector; determine the center of the road condition area image according to the center offset vector and the center of the image area corresponding to the interference target model in the image in front of the vehicle;

[0115] The rule for extracting the road condition area image is to map the moving direction to the image in front of the vehicle and determine the center offset vector. Specifically, the coordinates of the direction vector of the moving direction can be used to query the coordinate conversion table pre-configured corresponding to each image in front of the vehicle, and the query is converted into the two-dimensional coordinates in the image, so as to obtain the center offset vector; the direction vector corresponding to the moving direction in the coordinate conversion table is correspondingly associated with the center offset vector;

[0116] In addition, during extraction, the pre-configured area size determination table is queried according to the target detection results and moving speed corresponding to the interference target model to determine the size of the extracted road condition area image; in the area size determination table, the target detection results, moving speed and extraction size are correspondingly associated; in the area size determination table, the larger the size of the vehicle corresponding to the target detection result, the faster the moving speed, and the larger the extracted size range.

[0117] The present invention also provides a braking control system based on night vehicle target detection of improved KSC-YOLOV5, such as Figure 8 As shown, it includes: a space construction unit 1, a target detection unit 2, a distance measurement unit 3, a target construction unit 4 and a behavior analysis unit 5;

[0118] Among them, the space construction unit constructs a driving simulation space according to the current driving environment of the vehicle at night; the target detection unit inputs the front image of the vehicle taken in the current driving environment into the improved KSC-YOLOV5 model for target detection; the distance measurement unit controls the action of the distance measurement device according to the target detection result to obtain the distance measurement data; the target construction unit constructs the interference target model in the driving simulation space according to the target detection result and the distance measurement data; the behavior analysis unit analyzes the behavior of the interference target model, and issues a braking reminder and / or controls the vehicle braking according to the analysis result.

[0119] The space construction unit performs the following operations:

[0120] Retrieving the road parameters of the road where the vehicle is located from the road information storage base according to the positioning information of the vehicle;

[0121] According to the road parameters, the driving simulation space parameters and the initial model space template are retrieved from the simulation space determination library;

[0122] The driving simulation space is constructed according to the retrieved driving simulation space parameters and the initial model space template.

[0123] Among them, the control steps of the distance measuring device are as follows:

[0124] Using a preset feature extraction template, extract features from the target detection results and generate an analysis data set based on the extracted features;

[0125] The control parameters are determined based on the analysis data set and a pre-configured control parameter determination library.

[0126] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5, characterized in that: include: Construct a driving simulation space based on the current driving environment of the vehicle at night; Input the front image of the vehicle taken in the current driving environment into the improved KSC-YOLOV5 model for target detection; According to the target detection result, the distance measurement device is controlled to obtain distance measurement data; Based on the target detection results and distance measurement data, an interference target model is constructed in the driving simulation space; Analyze the behavior of the interference target model, and issue a braking reminder and / or control vehicle braking based on the analysis results; The behavior analysis steps include: According to the position change of the interference target model at the previous moment and the current moment, the moving direction of the interference target model is analyzed; Based on the interference target model and moving direction at the current moment, the trajectory of the interference target model is predicted; Determine whether the predicted trajectory intersects with the vehicle's trajectory, and output a braking reminder if an intersection occurs; Based on the preset trajectory, determine whether the interference target model will collide with the vehicle model corresponding to the vehicle, and if so, control the vehicle to brake; The behavioral analysis steps also include: According to the position change of the interference target model at the previous moment and the current moment, the moving direction and moving speed of the interference target model are analyzed as the moving parameters at the previous moment; According to the positioning of the vehicle and the relative position of the interference target model at each moment and the vehicle, the position of the interference target model at each moment in the pre-configured road condition data distribution map is determined and the corresponding road condition data is retrieved; the road condition data includes: road surface type and road surface flatness; Extract the road condition area image corresponding to the target corresponding to the interference target model in the front image of the vehicle at each time, input the pre-configured road condition analysis model, and obtain real-time road condition data; the real-time road condition data includes: whether there is an obstacle and obstacle size parameters; Construct an analysis data set for target stability analysis based on the movement parameters at each moment in the previous N moments, the corresponding retrieved road condition data, and the real-time road condition data; matching the analysis data set with a standard data set associated with each behavior analysis result in a pre-configured behavior analysis library; extracting behavioral analysis results associated with a standard data set that matches the analysis data set; Determine the behavior of other vehicles in the image in front of the vehicle from the behavior analysis result, and perform braking control when the behavior analysis result shows that the vehicle is out of control or has poor stability; The rule for extracting the road condition area image is to map the moving direction to the image in front of the vehicle and determine the center offset vector; determine the center of the road condition area image according to the center offset vector and the center of the image area corresponding to the interference target model in the image in front of the vehicle; Among them, the rule for extracting the road condition area image is to map the moving direction to the image in front of the vehicle, determine the center offset vector, use the coordinates of the direction vector of the moving direction, query the coordinate conversion table pre-configured corresponding to each image in front of the vehicle, and convert the query into a two-dimensional coordinate in the image to obtain the center offset vector; in the coordinate conversion table, the direction vector corresponding to the moving direction is correspondingly associated with the center offset vector.

2. The braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 1, characterized in that: The steps to construct the driving simulation space are as follows: Retrieving the road parameters of the road where the vehicle is located from the road information storage base according to the positioning information of the vehicle; According to the road parameters, the driving simulation space parameters and the initial model space template are retrieved from the simulation space determination library; The driving simulation space is constructed according to the retrieved driving simulation space parameters and the initial model space template.

3. The braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 1, characterized in that: The improved KSC-YOLOV5 model includes: SPPF module.

4. The braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 1, characterized in that: The Neck module of the improved KSC-YOLOV5 model includes: path aggregation network and feature pyramid network.

5. The braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 1, characterized in that: The improved KSC-YOLOV5 model adopts the SIOU loss function. The formula of the loss function is as follows: Where, SIOU loss represents the SIOU loss function; IOU represents the IOU loss; Δ is the distance loss under the condition of considering the angle loss; Ω represents the shape loss; B and B GT They represent the area of ​​the true bounding box and the area of ​​the predicted bounding box respectively.

6. The braking control method for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 1, characterized in that: The control steps of the distance measuring device are as follows: Using a preset feature extraction template, extract features from the target detection results and generate an analysis data set based on the extracted features; The control parameters are determined based on the analysis data set and a pre-configured control parameter determination library.

7. A braking control system for nighttime vehicle target detection based on improved KSC-YOLOV5, characterized in that: include: Space construction unit, target detection unit, distance measurement unit, target construction unit and behavior analysis unit; Among them, the space construction unit constructs a driving simulation space according to the current driving environment of the vehicle at night; the target detection unit inputs the front image of the vehicle taken in the current driving environment into the improved KSC-YOLOV5 model for target detection; the distance measurement unit controls the action of the distance measurement device according to the target detection result and obtains the distance measurement data; the target construction unit constructs the interference target model in the driving simulation space according to the target detection result and the distance measurement data; the behavior analysis unit analyzes the behavior of the interference target model, and issues a braking reminder and / or controls the vehicle braking according to the analysis result; The behavior analysis unit performs the following operations: According to the position change of the interference target model at the previous moment and the current moment, the moving direction of the interference target model is analyzed; Based on the interference target model and moving direction at the current moment, the trajectory of the interference target model is predicted; Determine whether the predicted trajectory intersects with the vehicle's trajectory, and output a braking reminder if an intersection occurs; Based on the preset trajectory, determine whether the interference target model will collide with the vehicle model corresponding to the vehicle, and if so, control the vehicle to brake; The behavior analysis unit also performs the following operations: According to the position change of the interference target model at the previous moment and the current moment, the moving direction and moving speed of the interference target model are analyzed as the moving parameters at the previous moment; According to the positioning of the vehicle and the relative position of the interference target model at each moment and the vehicle, the position of the interference target model at each moment in the pre-configured road condition data distribution map is determined and the corresponding road condition data is retrieved; the road condition data includes: road surface type and road surface flatness; Extract the road condition area image corresponding to the target corresponding to the interference target model in the front image of the vehicle at each time, input the pre-configured road condition analysis model, and obtain real-time road condition data; the real-time road condition data includes: whether there is an obstacle and obstacle size parameters; Construct an analysis data set for target stability analysis based on the movement parameters at each moment in the previous N moments, the corresponding retrieved road condition data, and the real-time road condition data; matching the analysis data set with a standard data set associated with each behavior analysis result in a pre-configured behavior analysis library; extracting behavioral analysis results associated with a standard data set that matches the analysis data set; Determine the behavior of other vehicles in the image in front of the vehicle from the behavior analysis result, and perform braking control when the behavior analysis result shows that the vehicle is out of control or has poor stability; The rule for extracting the road condition area image is to map the moving direction to the image in front of the vehicle and determine the center offset vector; determine the center of the road condition area image according to the center offset vector and the center of the image area corresponding to the interference target model in the image in front of the vehicle; Among them, the rule for extracting the road condition area image is to map the moving direction to the image in front of the vehicle, determine the center offset vector, use the coordinates of the direction vector of the moving direction, query the coordinate conversion table pre-configured corresponding to each image in front of the vehicle, and convert the query into a two-dimensional coordinate in the image to obtain the center offset vector; in the coordinate conversion table, the direction vector corresponding to the moving direction is correspondingly associated with the center offset vector.

8. The braking control system for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 7, characterized in that: The space construction unit performs the following operations: Retrieving the road parameters of the road where the vehicle is located from the road information storage base according to the positioning information of the vehicle; According to the road parameters, the driving simulation space parameters and the initial model space template are retrieved from the simulation space determination library; The driving simulation space is constructed according to the retrieved driving simulation space parameters and the initial model space template.

9. The braking control system for nighttime vehicle target detection based on improved KSC-YOLOV5 according to claim 7, characterized in that: The control steps of the distance measuring device are as follows: Using a preset feature extraction template, extract features from the target detection results and generate an analysis data set based on the extracted features; The control parameters are determined based on the analysis data set and a pre-configured control parameter determination library.

Images