Intelligent driving automobile point cloud target detection algorithm fused with SimAM

Abstract

The invention discloses an intelligent driving automobile point cloud target detection algorithm fused with SimAM. The algorithm comprises the following steps: acquiring three-dimensional point cloud data in a driving environment of an intelligent driving automobile; inputting the original three-dimensional point cloud data into the columnar feature extraction network and converting the original three-dimensional point cloud data into a pseudo image; performing feature extraction on the pseudo image by using a backbone network based on a SimAM attention mechanism and cavity convolution, and outputting feature maps with different sizes; respectively inputting the feature maps with different sizes into a multi-head SSD detector, and outputting category and position information of obstacles with different sizes; calculating total loss by using a loss function according to the category and position information; and performing iterative optimization on the process according to the total loss. The invention relates to the technical field of point cloud target detection, and solves the technical problems that in a traditional PointPillars method, due to the fact that the neuron receptive field is small and the feature extraction mode is simple, it is difficult to extract the features of a long-distance large obstacle, and the detection accuracy is low.

Description

Technical Field

[0001] The present invention belongs to the field of automobile point cloud target detection and relates to deep learning technology, and specifically is an intelligent driving automobile point cloud target detection algorithm integrated with SimAM. Background Art

[0002] With the rapid development of intelligent driving car technology, point cloud target detection technology, as an important part of environmental perception, has a decisive influence on the performance of the entire system in terms of accuracy and efficiency. The precise three-dimensional point cloud data generated by LiDAR provides vehicles with rich spatial information, but also brings challenges in data processing. The three-dimensional target detection algorithms based on laser point clouds are mainly divided into two categories: algorithms based on raw point clouds and algorithms based on voxels. The algorithm based on raw point clouds directly inputs the raw data into the neural network for training and outputs the detection results. Although there is no information loss and noise quantization problem, it has a large amount of calculation. The voxel-based algorithm converts the raw point cloud data through the voxel representation method, improves the data processing efficiency, and can process a large amount of point cloud data. For example, the PointPillars algorithm, as an efficient deep learning architecture, effectively reduces the computational complexity by dividing the continuous space into columnar structures and then using a set of convolutional neural networks to extract the features of each columnar structure.

[0003] However, the original PointPillars method has several significant problems: First, due to the limitation of the receptive field, it is difficult for the model to fully capture long-distance dependencies, especially when identifying large vehicle obstacles of 11 meters and above. Since the receptive field of neurons in the algorithm is small, the features of large obstacles cannot be fully extracted, resulting in the output of large obstacle sizes being smaller than the actual obstacle size; second, the simple feature extraction method causes the algorithm to not fully capture the features of complex or distant objects, thus affecting the accuracy of detection. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM, which is used to solve the technical problems in the traditional PointPillars method that it is difficult to extract features of large obstacles above 11 meters and the detection accuracy is low due to the small receptive field of neurons and the simple feature extraction method.

[0005] To achieve the above objectives, the present invention provides a point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM, comprising:

[0006] S1: Use laser radar to obtain three-dimensional point cloud data in the driving environment of the intelligent driving car to obtain original three-dimensional point cloud data;

[0007] S2: inputting the original three-dimensional point cloud data into a columnar feature extraction network and converting it into a pseudo image;

[0008] S3: extracting features from the pseudo image using a backbone network, and outputting feature maps of different sizes; wherein the backbone network is constructed based on the SimAM attention mechanism and dilated convolution;

[0009] S4: Input the feature maps of different sizes into the multi-head SSD detector respectively, and output the category and position information of obstacles at different sizes; finally, a car point cloud target detection model is obtained; wherein the position information is represented by a 3D box;

[0010] S5: According to the category and position information of the obstacle, the category loss, positioning loss and orientation loss are obtained using the loss function, and the total loss is calculated based on the category loss, positioning loss and orientation loss;

[0011] S6: Iteratively optimize the automobile point cloud target detection model according to the total loss to obtain an automobile point cloud target detection algorithm.

[0012] The improved PointPillars algorithm is obtained by converting the original 3D point cloud data into a pseudo image, using a backbone network that combines the SimAM attention mechanism and dilated convolution to perform deep feature extraction on the pseudo image, and then inputting the output feature maps of different sizes into the multi-head SSD detector. Among them, the SimAM attention mechanism can strengthen the features of key areas and suppress the noise in non-key areas by automatically learning and generating spatial attention maps, helping the model to better focus on important features; the dilated convolution expands the receptive field of the convolution kernel, capturing more contextual information without increasing the number of parameters, and effectively solves the problem of small receptive field in the original PointPillars algorithm. This improvement enables the model to more fully extract the features of large obstacles and improve the detection accuracy of complex or distant objects.

[0013] Furthermore, the original three-dimensional point cloud data is input into a columnar feature extraction network to be converted into a pseudo image, comprising the following steps:

[0014] A1, establish an xy plane, divide the point cloud samples in the original three-dimensional point cloud data into n×m grids on the xy plane, generate a point column for each grid in the z-axis direction, and obtain a number of point columns; wherein the points in the point columns are represented by a nine-dimensional vector j: (x, y, z, r, x c ,y c ,z c ,x p ,y p ), where x, y, and z represent the position of the point in space coordinates, r represents the laser radar reflection intensity, and xc ,y c 、z c Indicates the geometric center of all points in the grid where the point is located, x p ,y p Represents the offset of each point in the point column relative to the geometric center of each grid; where n and m represent the height and width of the grid respectively, and both are positive integers greater than 0;

[0015] A2, compare the number of point clouds in a number of point columns with the point count threshold d; if the number of point clouds is greater than the point count threshold d, randomly sample d points and keep them in the point column; if the number of point clouds is less than the point count threshold d, assign 0 to the nine-dimensional vector j to fill the number of point clouds to d;

[0016] A3, encoding the plurality of point cloud samples into a tensor of (j, p, d); wherein p represents the number of point columns of each point cloud sample;

[0017] A4, using a 1×1 convolution kernel to perform a convolution operation on the points in the plurality of point columns, through BatchNorm, Relu function and performing a maxpooling operation on the channel, outputting a tensor of size (f, p); projecting the tensor of size (f, p) into a tensor of size (f, h, w) to obtain a pseudo image; wherein f is the number of channels of the pseudo image, and h and w are the height and width of the pseudo image.

[0018] In the process of converting the original point cloud into a pseudo image, the amount of calculation is reduced by setting the number of point columns and randomly sampling or filling the points in the point cloud, while the ability of the pseudo image to express the original point cloud information is improved.

[0019] Furthermore, the backbone network includes: a bottom-up downsampling network, a top-down upsampling network and a SimAM attention module; the downsampling network includes two downsampling operations, the first downsampling uses a number of 3×3 convolutions with a step size of s, and the second downsampling uses a number of 3×3 dilated convolutions with a dilation rate of r;

[0020] The upsampling network uses the nearest neighbor interpolation algorithm to perform upsampling twice step by step, and performs a feature fusion operation on the feature map in the upsampling process and the feature map of the same size in the downsampling process to obtain feature maps of different sizes after feature fusion;

[0021] The SimAM attention module is used to perform feature processing on the feature maps of different sizes after the feature fusion to generate new feature maps of different sizes.

[0022] Furthermore, the calculation steps of the SimAM attention module are as follows:

[0023] B1, define the energy function for each neuron in the neural network: Among them, t and x i They represent the target neuron and other neurons of the input feature X, i is the spatial dimension index, M is the number of neurons on a specific channel, and w t , b t is the transformation weight and bias of the neuron, λ is the regularization coefficient;

[0024] B2, according to the formula Derive the analytical solution; where μ t , σ 2 The calculation formula is

[0025] B3, through the formula Calculate the minimum energy; where The calculation formula is

[0026] B4, applying the minimum energy, according to the formula Enhance feature X to obtain the feature processed by the attention module Where ⊙ represents matrix dot product and sigmoid() represents the activation function.

[0027] The introduction of dilated convolution can effectively increase the size of the feature map without increasing the network depth, the number of parameters and the computational complexity, so that the model can capture richer contextual information; SimAM accelerates the calculation of attention weights by deriving the analytical solution of the energy function, which enables SimAM to have higher computational efficiency while further strengthening the key features; the organic combination of the two enables the model to more accurately extract the features of large obstacles, thereby solving the problem of small output size of large obstacles in the traditional PointPillars algorithm and improving the detection accuracy of complex or distant objects.

[0028] Furthermore, the category loss is obtained by using the calculation formula of Weight softmax loss: Get the category loss; where p i 、p j Represent the output values ​​of the i-th and j-th nodes respectively, C is the number of output nodes, and w i 、w j denote the weights of the i-th and j-th output nodes respectively.

[0029] Furthermore, the positioning loss is obtained by using the calculation formula of Smooth L1 loss: Get the positioning loss of a single category and use the formula L loc=∑ bε(x,y,z,w,l,h,θ) SmoothL1(Δb) calculates the overall positioning loss; where x', y' represent the output and label of the model, x, y, z represent the center coordinates of the 3D box, w, l, h represent the length, width and height of the 3D box, θ represents the orientation of the 3D box, and (x, y, z, w, l, h, θ) represents each 3D box vector.

[0030] Furthermore, the orientation loss is obtained by using the calculation formula L of Softmax classification loss: dir =SmoothL1(sin(θ p -θ q )) get the direction loss; where θ p is the predicted angle, θ q The real angle.

[0031] Furthermore, the calculation formula of the total loss is: Among them, N pos is the number of predefined positive anchor boxes.

[0032] The category loss encourages the model to output Gaussian distribution results, so that the network tends to predict the most likely category, which helps to improve classification accuracy. The positioning loss can reduce problems caused by gradient disappearance or explosion, help the model better learn the location information of the bounding box, and make the model more accurate in locating the bounding box. The orientation loss emphasizes the angle differences between different categories, so that the model can better understand the angle characteristics unique to each category.

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

[0034] By integrating the SimAM attention mechanism and using the dilated convolution technology, the detection accuracy and performance of the intelligent driving car point cloud target detection technology based on the PointPillars algorithm are improved. By adding the SimAM attention mechanism after feature fusion, the features of the key areas are strengthened and the noise of the non-key areas is suppressed, so that the model can focus on important features more accurately and improve the detection accuracy of complex or distant objects; and the dilated convolution technology is used in the second downsampling of the backbone network, which effectively increases the size of the feature map without increasing the network depth and the number of parameters, captures richer context information, and solves the problem of small receptive field in the traditional PointPillars algorithm; the improved PointPillars algorithm outputs feature maps of different sizes through the backbone network, and uses a multi-head SSD detector to detect and process obstacles, which optimizes the model's ability to recognize large obstacles and avoids the problem of small output result size; through the comprehensive application of category loss, positioning loss and orientation loss, the classification accuracy and the positioning accuracy of the bounding box are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] 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 of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 A flow chart of a point cloud target detection algorithm for an intelligent driving vehicle integrated with SimAM provided by the present invention;

[0037] Figure 2 A framework diagram of a point cloud target detection algorithm model for an intelligent driving vehicle integrated with SimAM provided by the present invention;

[0038] Figure 3 A schematic diagram of a columnar feature extraction network framework for converting three-dimensional point cloud data into a pseudo image provided by the present invention;

[0039] Figure 4 Schematic diagram of the backbone network framework combining dilated convolution and SimAM provided by the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than 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.

[0041] See also Figure 1-Figure 4 , an embodiment of the present invention provides a point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM, including:

[0042] S1: Use laser radar to obtain three-dimensional point cloud data in the driving environment of the intelligent driving car to obtain original three-dimensional point cloud data;

[0043] S1: Use laser radar to obtain three-dimensional point cloud data in the driving environment of the intelligent driving car to obtain original three-dimensional point cloud data;

[0044] S2: inputting the original three-dimensional point cloud data into a columnar feature extraction network and converting it into a pseudo image;

[0045] S3: extracting features from the pseudo image using a backbone network, and outputting feature maps of different sizes; wherein the backbone network is constructed based on the SimAM attention mechanism and dilated convolution;

[0046] S4: Input the feature maps of different sizes into the multi-head SSD detector respectively, and output the category and position information of obstacles at different sizes; finally, a car point cloud target detection model is obtained; wherein the position information is represented by a 3D box;

[0047] S5: According to the category and position information of the obstacle, the category loss, positioning loss and orientation loss are obtained using the loss function, and the total loss is calculated based on the category loss, positioning loss and orientation loss;

[0048] S6: Iteratively optimize the automobile point cloud target detection model according to the total loss to obtain an automobile point cloud target detection algorithm.

[0049] In this embodiment, the specific operation method of converting the original three-dimensional point cloud data into a pseudo image is:

[0050] First, an xy plane is established. Several point cloud samples in the original 3D point cloud data are divided into h×w grids on the xy plane. A point column is generated for each grid in the z-axis direction to obtain several point columns. Among them, the points in the several point columns are represented by nine-dimensional vectors j: (x, y, z, r, x c ,y c ,z c ,x p ,y p), where x, y, and z represent the position of the point in space coordinates, r represents the laser radar reflection intensity, and x c ,y c 、z c Indicates the geometric center of all points in the grid where the point is located, x p ,y p Represents the offset of each point in the point column relative to the geometric center of each grid; where h and w are both positive integers greater than 0;

[0051] Then, the number of point clouds in several point columns is compared with the point threshold d. Assuming that the point threshold d is 100, if the number of point clouds is greater than 100, 100 points are randomly sampled and retained in the point column; if the number of point clouds is less than 100, 0 is assigned to the nine-dimensional vector j to fill the number of point clouds to 100;

[0052] Next, several point cloud samples are encoded into a tensor of (j, p, d); where p represents the number of point columns of each point cloud sample;

[0053] Finally, a 1×1 convolution kernel is used to perform a convolution operation on the points in several point columns, and a tensor of size (f, p) is output through BatchNorm, Relu function, and maxpooling operation on the channel; the tensor of size (f, p) is projected into a tensor of size (f, h, w) to obtain a pseudo image with f channels, h height, and w width.

[0054] In the original PointPillars algorithm, after the point cloud data is converted into a pseudo image, a simple two-dimensional convolutional neural network is used as the backbone network to extract features from the pseudo image. The feature extraction ability of this backbone network is limited by the size of the receptive field, making it difficult for the model to capture global information. Therefore, the original backbone network is improved by combining the dilated convolution and the SimAM attention mechanism;

[0055] SimAM is a lightweight, parameter-free attention mechanism inspired by neuroscience that enhances feature representation by evaluating the importance of each neuron. Specifically, the calculation steps of SimAM are as follows:

[0056] First, define the energy function for each neuron in the neural network: Among them, t and x i They represent the target neuron and other neurons of the input feature X, i is the spatial dimension index, M is the number of neurons on a specific channel, and w t , b t is the transformation weight and bias of the neuron, λ is the regularization coefficient;

[0057] Then, according to the formula Derive the analytical solution; where μ t , σ 2 The calculation formula is

[0058] Then, through the formula Calculate the minimum energy; where The calculation formula is

[0059] Finally, the minimum energy is applied, according to the formula Enhance feature X to obtain the feature processed by the attention module where ⊙ represents matrix dot product.

[0060] Based on the above calculation steps, the improved backbone network includes: bottom-up downsampling network, top-down upsampling network and SimAM attention module;

[0061] First, before downsampling, the pseudo image is convolved with a kernel of 3×3 to obtain a feature map of size (2f, h / 2, w / 2); then it enters the downsampling network;

[0062] The downsampling network includes two downsampling operations. In the first downsampling, several 3×3 convolutions with a step size of s are used to obtain a feature map of size (4f, h / 4, w / 4); in the second downsampling, several 3×3 dilated convolutions with a dilation rate of r are used to obtain a feature map of size (8f, h / 8, w / 8);

[0063] The upsampling network uses the nearest neighbor interpolation algorithm to perform two level-by-level upsampling, and then performs feature fusion operation on the feature map in the upsampling process and the feature map of the same size in the downsampling process to obtain feature maps of different sizes after feature fusion;

[0064] The SimAM attention module is used to process the feature maps of different sizes after feature fusion to generate new feature maps of different sizes, namely (8f, h / 8, w / 8), (4f, h / 4, w / 4), and (2f, h / 2, w / 2).

[0065] Finally, after performing 3×3 convolution operations on the fused feature maps of different sizes, the output feature maps are obtained for the input of the multi-head SSD detector.

[0066] After the improved backbone network extracts features from the pseudo-image, it will output feature maps of different sizes, which are then input into the multi-head SSD detector to predict the category and location information of the object on the feature maps of different scales. Finally, the model is continuously iterated and optimized through the loss function:

[0067] The loss function is a key indicator in the model optimization process, which measures the difference between the model prediction results and the actual labels. By minimizing the loss function, the model can gradually adjust its parameters to better fit the training data, thereby improving the generalization ability on unseen data. In this embodiment, Weight softmax loss, Smooth L1loss, and Softmax classification loss are used as the category loss, positioning loss, and orientation loss of the model;

[0068] Specifically, using the calculation formula of Weight softmax loss Get the category loss; where p i 、p j Represent the output values ​​of the i-th and j-th nodes respectively, C is the number of output nodes, and w i 、w j denote the weights of the i-th and j-th output nodes respectively.

[0069] Using the calculation formula of Smooth L1 loss Get the positioning loss of a single category and use the formula L loc =∑ bε(x,y,z,w,l,h,θ) SmoothL1(Δb) calculates the overall positioning loss; where x', y' represent the output and label of the model, x, y, z represent the center coordinates of the 3D box, w, l, h represent the length, width and height of the 3D box, θ represents the orientation of the 3D box, and (x, y, z, w, l, h, θ) represents each 3D box vector.

[0070] Using the calculation formula L of Softmax classification loss dir =SmoothL1(sin(θ p -θ q )) get the direction loss; where θ p is the predicted angle, θ q is the real angle;

[0071] Finally, by weighting, the calculation formula for the total loss is obtained: Among them, N pos is the number of predefined positive anchor boxes.

[0072] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0073] Working principle of the present invention:

[0074] S1, data acquisition and conversion: using LiDAR to collect 3D point cloud data in the vehicle driving environment; then, using a columnar feature extraction network to convert the 3D point cloud data into a pseudo image for subsequent processing;

[0075] S2, feature extraction: The backbone network is used to perform deep feature extraction on the pseudo image to obtain feature maps of different sizes. The backbone network combines the SimAM attention mechanism and dilated convolution to increase the receptive field, effectively capture global information, and strengthen the features of key areas while suppressing noise in non-key areas;

[0076] S3, object detection: The extracted feature maps of different sizes are input into the multi-head SSD detector, which can predict the category and location information of objects at different scales;

[0077] S4, loss calculation and model optimization: Based on the category and location information of the detected obstacles, the category loss, positioning loss and orientation loss are calculated, and the total loss is obtained by weighted combination. Finally, the model is iteratively optimized according to the total loss to improve detection accuracy and reduce errors.

[0078] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM, characterized in that: include: S1: Use laser radar to obtain three-dimensional point cloud data in the driving environment of the intelligent driving car to obtain original three-dimensional point cloud data; S2: inputting the original three-dimensional point cloud data into a columnar feature extraction network and converting it into a pseudo image; S3: extracting features from the pseudo image using a backbone network, and outputting feature maps of different sizes; wherein the backbone network is constructed based on the SimAM attention mechanism and dilated convolution; S4: Input the feature maps of different sizes into the multi-head SSD detector respectively, and output the category and position information of obstacles at different sizes; finally, a car point cloud target detection model is obtained; wherein the position information is represented by a 3D box; S5: According to the category and position information of the obstacle, the category loss, positioning loss and orientation loss are obtained by using the loss function, and the total loss is calculated based on the category loss, positioning loss and orientation loss; S6: Iteratively optimize the automobile point cloud target detection model according to the total loss to obtain an automobile point cloud target detection algorithm.

2. According to claim 1, the point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM is characterized in that: Inputting the original three-dimensional point cloud data into a columnar feature extraction network to convert it into a pseudo image comprises the following steps: A1, establish an xy plane, divide the point cloud samples in the original three-dimensional point cloud data into n×m grids on the xy plane, generate a point column for each grid in the z-axis direction, and obtain a number of point columns; wherein the points in the point columns are represented by a nine-dimensional vector j: (x, y, z, r, x c ,y c ,z c ,x p ,y p ); x, y, z represent the position of the point in the spatial coordinates, r represents the laser radar reflection intensity, x c ,y c 、z c Indicates the geometric center of all points in the grid where the point is located, x p ,y p Represents the offset of each point in the point column relative to the geometric center of each grid; n and m represent the height and width of the grid respectively, and both are positive integers greater than 0; A2, compare the number of point clouds in a number of point columns with the point count threshold d; if the number of point clouds is greater than the point count threshold d, randomly sample d points and keep them in the point column; if the number of point clouds is less than the point count threshold d, assign 0 to the nine-dimensional vector j to fill the number of point clouds to d; A3, encoding the plurality of point cloud samples into a tensor of (j, p, d); wherein p represents the number of point columns of each point cloud sample; A4, using a 1×1 convolution kernel to perform a convolution operation on the points in the plurality of point columns, through BatchNorm, Relu function and performing a maxpooling operation on the channel, outputting a tensor of size (f, p); projecting the tensor of size (f, p) into a tensor of size (f, h, w) to obtain a pseudo image; wherein f is the number of channels of the pseudo image, and h and w are the height and width of the pseudo image.

3. The point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM according to claim 1 is characterized in that: The backbone network includes: a bottom-up downsampling network, a top-down upsampling network and a SimAM attention module; the downsampling network includes two downsampling operations, the first downsampling uses a number of 3×3 convolutions with a step size of s, and the second downsampling uses a number of 3×3 dilated convolutions with a dilation rate of r; The upsampling network uses the nearest neighbor interpolation algorithm to perform upsampling twice step by step, and performs a feature fusion operation on the feature map in the upsampling process and the feature map of the same size in the downsampling process to obtain feature maps of different sizes after feature fusion; The SimAM attention module is used to perform feature processing on the feature maps of different sizes after the feature fusion to generate new feature maps of different sizes.

4. The point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM according to claim 3 is characterized in that: The calculation steps of the SimAM attention module are as follows: B1, define the energy function for each neuron in the neural network: Among them, t and x i They represent the target neuron and other neurons of the input feature X, i is the spatial dimension index, M is the number of neurons on a specific channel, and w t 、b t is the transformation weight and bias of the neuron, λ is the regularization coefficient; B2, according to the formula Derive the analytical solution; where μ t , σ 2 The calculation formula is B3, through the formula Calculate the minimum energy; where The calculation formula is B4, applying the minimum energy, according to the formula Enhance feature X to obtain the feature processed by the attention module Where ⊙ represents matrix dot product and sigmoid() represents the activation function.

5. The point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM according to claim 1 is characterized in that: The category loss is obtained by using the calculation formula of Weight softmax loss Get the category loss; Among them, p i 、p j Represent the output values ​​of the i-th and j-th nodes respectively, C is the number of output nodes, and w i 、w j denote the weights of the i-th and j-th output nodes respectively.

6. The point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM according to claim 1 is characterized in that: The positioning loss is obtained by using the calculation formula of Smooth L1 loss: Get the positioning loss of a single category and use the formula L loc =∑ bε(x,y,z,w,l,h,θ) SmoothL1(Δb) calculates the overall positioning loss; where x', y' represent the output and label of the model, x, y, z represent the center coordinates of the 3D box, w, l, h represent the length, width and height of the 3D box, θ represents the orientation of the 3D box, and (x, y, z, w, l, h, θ) represents each 3D box vector.

7. The point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM according to claim 1 is characterized in that: The orientation loss is obtained by using the calculation formula L of Softmax classification loss dir =SmoothL1(sin(θ p -θ q )) get the direction loss; where θp 为 Prediction angle, θ q The real angle.

8. The point cloud target detection algorithm for intelligent driving vehicles integrated with SimAM according to claim 1 is characterized in that: The total loss is calculated as follows: Among them, N pos is the number of predefined positive anchor boxes.

Images