Training method and device of target detection model and target detection method and device

Abstract

The application discloses a target detection model training method and device, and a target detection method and device. The target detection model training method comprises the following steps: acquiring laser point cloud data for training, processing the laser point cloud data by using a point cloud feature processing network in the target detection model, and obtaining processed point cloud features; extracting features of the processed point cloud features by using a point cloud feature extraction network in the target detection model, and obtaining a point cloud feature extraction result, wherein the point cloud feature extraction network is obtained by adding a convolution layer in a low layer scale space in an original FPN network; performing target detection on the point cloud feature extraction result by using a detection head network in the target detection model, and obtaining a target detection result; and determining a loss value of the target detection result by using a preset loss function, so as to update parameters of the target detection model. The application enhances the extraction capability of the original target detection model for low layer semantic information, and improves the detection precision of the model for small targets.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a target detection model training method and device and a target detection method and device. BACKGROUND

[0002] With the rapid development of the intelligent driving field, vehicles are joining more and more driver assistance functions, such as a lane keeping function, which allows the vehicle to be positioned correctly within the lane of the road, which is also crucial for subsequent lane deviation and trajectory planning decisions of fully autonomous vehicles.

[0003] In the field of driving assistance or autonomous driving, target detection is crucial for the realization of intelligent driving functions such as road planning and lane keeping of vehicles. Existing target detection methods rely on deep learning algorithms, which are usually followed by post-processing techniques. This brings a high computational cost and poor model scalability due to changes in road scenes. In addition, existing end-to-end deep learning algorithms have poor detection effect for small targets when performing target detection, resulting in low detection accuracy. Post-processing algorithms are needed to optimize them, but such methods also have poor scalability. SUMMARY

[0004] The embodiments of the present application provide a target detection model training method and device and a target detection method and device to improve target detection accuracy.

[0005] The embodiments of the present application adopt the following technical solutions:

[0006] In a first aspect, the embodiments of the present application provide a target detection model training method, which includes:

[0007] Obtaining laser point cloud data for training, and processing the laser point cloud data using a point cloud feature processing network in the target detection model to obtain processed point cloud features;

[0008] Using a point cloud feature extraction network in the target detection model to extract features from the processed point cloud features to obtain point cloud feature extraction results, the point cloud feature extraction network being obtained by adding a convolution layer in a low layer scale space in an original feature pyramid network;

[0009] Using a detection head network in the target detection model to perform target detection on the point cloud feature extraction results to obtain target detection results;

[0010] Using a preset loss function to determine a loss value of the target detection results, and updating parameters of the target detection model according to the loss value of the target detection results to obtain a trained target detection model.

[0011] Optionally, the processing of the laser point cloud data by using a point cloud feature processing network in the target detection model comprises:

[0012] The point cloud space corresponding to the laser point cloud data is divided to obtain a plurality of columnar bodies of the same size;

[0013] The points in each columnar body are sampled and processed by using a farthest point sampling algorithm to obtain a plurality of columnar bodies after sampling and processing;

[0014] The laser point cloud data corresponding to the plurality of columnar bodies after sampling and processing is processed by using a point cloud feature processing network to obtain the processed point cloud features.

[0015] Optionally, the processing of the laser point cloud data by using a point cloud feature processing network in the target detection model comprises:

[0016] The point cloud space corresponding to the laser point cloud data is divided to obtain a plurality of columnar bodies of the same size;

[0017] A plurality of point cloud feature extraction strategies are used to extract point cloud features of each columnar body to obtain a plurality of point cloud features corresponding to each columnar body;

[0018] The plurality of point cloud features corresponding to each columnar body are spliced respectively to obtain the processed point cloud features corresponding to each columnar body.

[0019] Optionally, the plurality of point cloud features comprise a first point cloud feature, a second point cloud feature and a third point cloud feature, and the processing of the laser point cloud data by using a plurality of point cloud feature extraction strategies to extract point cloud features of each columnar body to obtain a plurality of point cloud features corresponding to each columnar body comprises:

[0020] The points in each columnar body are dimensioned by using a full connection layer in the point cloud feature processing network, and the number of points in each columnar body is compressed by using a maximum pooling layer in the point cloud feature processing network to obtain the first point cloud feature corresponding to each columnar body;

[0021] The points in each columnar body are spliced into a matrix, and a first convolution layer in the point cloud feature processing network is used to convolve the matrix corresponding to each columnar body to obtain the second point cloud feature corresponding to each columnar body;

[0022] A second convolution layer in the point cloud feature processing network is used to convolve the multi-dimensional features corresponding to each point in each columnar body to obtain the third point cloud feature corresponding to each columnar body.

[0023] Optionally, the original feature pyramid network comprises a low-level scale convolutional layer and a corresponding de-convolutional layer, and a high-level scale convolutional layer and a corresponding de-convolutional layer, the feature extraction on the processed point cloud feature by using the point cloud feature extraction network in the target detection model to obtain a point cloud feature extraction result comprises:

[0024] A preset convolutional layer is added between the low-level scale convolutional layer and the corresponding de-convolutional layer;

[0025] The processed point cloud feature is convoluted by using the low-level scale convolutional layer to obtain a first convolutional result of the low-level scale convolutional layer;

[0026] The first convolutional result of the low-level scale convolutional layer is convoluted by using the preset convolutional layer to obtain a second convolutional result of the low-level scale convolutional layer;

[0027] The first convolutional result of the low-level scale convolutional layer is convoluted by using the high-level scale convolutional layer to obtain a convolutional result of the high-level scale convolutional layer;

[0028] The second convolutional result of the low-level scale convolutional layer and the convolutional result of the high-level scale convolutional layer are spliced to obtain the point cloud feature extraction result.

[0029] Optionally, the original feature pyramid network comprises a first convolutional layer, a second convolutional layer and a third convolutional layer, the feature extraction on the processed point cloud feature by using the point cloud feature extraction network in the target detection model to obtain a point cloud feature extraction result comprises:

[0030] A first maximum pooling layer is added between the first convolutional layer and the second convolutional layer, and a second maximum pooling layer is added between the second convolutional layer and the third convolutional layer;

[0031] The processed point cloud feature is convoluted by using the first convolutional layer to obtain a convolutional result of the first convolutional layer;

[0032] The convolutional result of the first convolutional layer is pooled by using the maximum pooling layer between the first convolutional layer and the second convolutional layer to obtain a pooling result of the first maximum pooling layer;

[0033] The pooling result of the first maximum pooling layer is convoluted by using the second convolutional layer to obtain a convolutional result of the second convolutional layer;

[0034] The convolutional result of the second convolutional layer is pooled by using the maximum pooling layer between the second convolutional layer and the third convolutional layer to obtain a pooling result of the second maximum pooling layer;

[0035] The pooling result of the second max pooling layer is convolved by the third convolutional layer to obtain the convolution result of the third convolutional layer;

[0036] The convolution results of the first convolutional layer, the second convolutional layer, and the third convolutional layer are deconvolved and then concatenated to obtain the point cloud feature extraction result.

[0037] Optionally, the detection head network includes multiple detection heads, and the step of using the detection head network in the target detection model to perform target detection on the point cloud feature extraction results to obtain target detection results includes:

[0038] The point cloud feature extraction results are detected by multiple detection heads respectively to obtain the detection results of multiple detection heads, and the detection results of multiple detection heads are stitched together to obtain the stitched detection result;

[0039] The concatenated detection results are predicted by multiple prediction branches in the detection head network to obtain multiple prediction results, which are used as the target detection results. The multiple prediction results include category prediction results, bounding box prediction results, orientation prediction results, and foreground / background prediction results.

[0040] The step of determining the loss value of the target detection result using a preset loss function and updating the parameters of the target detection model based on the loss value of the target detection result includes:

[0041] The category prediction loss is determined based on the category prediction result, the box prediction loss is determined based on the box prediction result, the direction prediction loss is determined based on the direction prediction result, and the foreground / background prediction loss is determined based on the foreground / background prediction result.

[0042] The parameters of the object detection model are updated based on the category prediction loss, the bounding box prediction loss, the orientation prediction loss, and the foreground / background prediction loss.

[0043] Secondly, embodiments of this application also provide a target detection method, wherein the target detection method includes:

[0044] Obtain the current laser point cloud data;

[0045] The target detection model is used to perform target detection on the current laser point cloud data to obtain the current target detection results;

[0046] The target detection model is trained based on any of the aforementioned target detection model training methods.

[0047] Thirdly, embodiments of this application also provide a training apparatus for an object detection model, wherein the training apparatus for the object detection model includes:

[0048] The point cloud feature processing unit is used to acquire laser point cloud data for training and to process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features.

[0049] The point cloud feature extraction unit is used to extract features from the processed point cloud features using the point cloud feature extraction network in the target detection model, and obtain the point cloud feature extraction result. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network.

[0050] The first detection unit is used to perform target detection on the point cloud feature extraction results using the detection head network in the target detection model, and obtain the target detection result.

[0051] The update unit is used to determine the loss value of the target detection result using a preset loss function, and update the parameters of the target detection model according to the loss value of the target detection result to obtain the trained target detection model.

[0052] Fourthly, embodiments of this application also provide a target detection device, wherein the target detection device includes:

[0053] The acquisition unit is used to acquire the current laser point cloud data;

[0054] The second detection unit is used to perform target detection on the current laser point cloud data using the target detection model, and obtain the current target detection result.

[0055] The target detection model is trained based on the training device for the aforementioned target detection model.

[0056] Fifthly, embodiments of this application also provide an electronic device, including:

[0057] Processor; and

[0058] A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform either the training method of any of the aforementioned target detection models or to perform the aforementioned target detection method.

[0059] Sixthly, embodiments of this application also provide a computer-readable storage medium storing one or more programs, which, when executed by an electronic device including multiple applications, cause the electronic device to perform a training method for any of the aforementioned target detection models, or to perform the aforementioned target detection method.

[0060] The above-mentioned at least one technical solution adopted in the embodiments of this application can achieve the following beneficial effects: The training method of the target detection model in the embodiments of this application first acquires laser point cloud data for training, and processes the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features; then, it uses the point cloud feature extraction network in the target detection model to extract features from the processed point cloud features to obtain point cloud feature extraction results. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network; then, it uses the detection head network in the target detection model to perform target detection on the point cloud feature extraction results to obtain target detection results; finally, it uses a preset loss function to determine the loss value of the target detection results, and updates the parameters of the target detection model according to the loss value of the target detection results to obtain the trained target detection model. The training method of the target detection model in the embodiments of this application improves the original PointPillars network, enhances the ability to extract semantic information in the low-level scale space, reduces the noise contained in the feature map of the low-level scale space, and thus improves the detection accuracy of the target detection model for small targets. Attached Figure Description

[0061] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0062] Figure 1 This is a flowchart illustrating a training method for an object detection model according to an embodiment of this application;

[0063] Figure 2 This is a schematic diagram of a primitive feature pyramid network in the prior art.

[0064] Figure 3 This is a schematic diagram of an improved feature pyramid network in an embodiment of this application;

[0065] Figure 4 This is a schematic diagram of another improved feature pyramid network in the embodiments of this application;

[0066] Figure 5 This is a schematic diagram of the original detection head network in a PointPillars network in the prior art;

[0067] Figure 6 This is a schematic diagram of an improved detection head network in an embodiment of this application;

[0068] Figure 7 This is a schematic diagram of the structure of a training device for an object detection model according to an embodiment of this application;

[0069] Figure 8 This is a flowchart illustrating a target detection method in an embodiment of this application;

[0070] Figure 9 This is a schematic diagram of the structure of a target detection device according to an embodiment of this application;

[0071] Figure 10 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0073] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0074] To facilitate understanding of the various embodiments of this application, we first introduce an existing target detection network, PointPillars. PointPillars is a fast target detection network based on point clouds, which achieves a good balance between speed and accuracy. It has wide applications, especially in the field of autonomous driving, where it has significant application value.

[0075] The overall network structure of the PointPillars network mainly includes a Pillar FeatureNet (PFN), a Feature Pyramid Network (FPN), and a 3D object detection head network. The Pillar FeatureNet is mainly used for spatial partitioning of point cloud data and 3D-to-2D conversion. The Feature Pyramid Network, as the backbone network (Backbone 2DCNN), is used to extract point cloud features from the 2D processing results output by the Pillar FeatureNet. The 3D object detection head network (Detection Head) is based on the SSD (Single Shot Detection) network and is used to predict the point cloud feature extraction results output by the Feature Pyramid Network, thereby obtaining the object detection results.

[0076] While the PointPillars network exhibits good detection performance, its effectiveness in detecting small targets still needs improvement. Therefore, the target detection model trained in this embodiment improves the structure of the original PointPillars network to further enhance target detection performance, particularly the accuracy of small target detection.

[0077] This application provides a method for training an object detection model, such as... Figure 1 The diagram shows a flowchart of a training method for an object detection model according to an embodiment of this application. The training method for the object detection model includes at least the following steps S110 to S140:

[0078] Step S110: Obtain laser point cloud data for training, and process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features.

[0079] The target detection model trained in this application embodiment is mainly based on the point cloud data of LiDAR for 3D target detection. Therefore, it is necessary to first obtain the LiDAR point cloud data for training. The LiDAR point cloud data can be collected in advance by a vehicle equipped with LiDAR and the data is labeled.

[0080] After obtaining the laser point cloud data for training, the laser point cloud data can be processed using the point cloud feature processing network in the target detection model to obtain the processed point cloud features. The point cloud feature processing network can be based on the point cloud feature processing network in the PointPillars network.

[0081] Step S120: Use the point cloud feature extraction network in the target detection model to extract features from the processed point cloud features to obtain point cloud feature extraction results. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network.

[0082] Then, the point cloud feature extraction network in the object detection model is used to extract features from the processed point cloud features. The point cloud feature extraction network here is obtained by modifying the original feature pyramid network in the PointPillars network.

[0083] Since the information of small targets is mainly stored in the low-scale feature space, but the original feature pyramid network in the PointPillars network has fewer convolutional layers in the low-scale feature space, the resulting feature map contains a lot of noise information. Therefore, this embodiment further adds convolutional layers in the low-scale space of the original feature pyramid network to enhance the extraction of feature semantic information in the low-scale space, reduce noise in the feature map, and thus improve the detection results of small targets.

[0084] Step S130: Use the detection head network in the target detection model to perform target detection on the point cloud feature extraction results to obtain the target detection results.

[0085] After obtaining the point cloud feature extraction results, the detection head network in the object detection model can be used to predict the point cloud feature extraction results. Specifically, this can include category prediction, bounding box position prediction, and orientation prediction, thereby obtaining the object detection results. The detection head network here can also be based on the 3D object detection head network in the PointPillars network.

[0086] Step S140: Determine the loss value of the target detection result using a preset loss function, and update the parameters of the target detection model according to the loss value of the target detection result to obtain the trained target detection model.

[0087] Based on the different dimensions of the target detection results and the labeled information in the laser point cloud data, the loss values ​​of the target detection results can be calculated using predefined loss functions. For example, the Focal Loss function can be used to calculate the loss for category prediction, the Smooth L1 function can be used to calculate the loss for bounding box position prediction, and the cross-entropy loss function can be used to calculate the loss for orientation prediction. Finally, the total loss value is calculated by combining the loss values ​​of each dimension and their corresponding loss weights, and the parameters of the target detection model are updated accordingly to obtain the trained target detection model.

[0088] The training method of the target detection model in this application improves the original PointPillars network, enhances the ability to extract semantic information in the low-scale space, reduces the noise contained in the feature map of the low-scale space, and thus improves the detection accuracy of the target detection model for small targets.

[0089] In some embodiments of this application, the step of processing the laser point cloud data using a point cloud feature processing network in the target detection model to obtain processed point cloud features includes: dividing the point cloud space corresponding to the laser point cloud data to obtain multiple columns of the same size; sampling points in each column using a farthest point sampling algorithm to obtain multiple columns after sampling processing; and performing point cloud feature processing on the laser point cloud data corresponding to the multiple columns after sampling processing to obtain the processed point cloud features.

[0090] Since the original point cloud feature processing network in the PointPillars network first divides the laser point cloud data into multiple columnar pillars of the same size, in order to ensure that each pillar contains N point cloud data, the points in each pillar are sampled through random sampling operations. For example, for pillars with more than N points, N points are randomly sampled, and for pillars with less than N points, N points are randomly filled in.

[0091] However, the aforementioned random sampling method only ensures that the number of points in each pillar is the same, but the distribution of points in each pillar is often uneven, which can affect the subsequent feature extraction and detection results. Therefore, this application improves the original point cloud feature processing network in the PointPillars network by using the Farthest Point Sampling (FPS) algorithm to sample each pillar separately, thereby ensuring a more uniform distribution of points in each pillar. After obtaining multiple pillars after sampling, point cloud feature processing can be further performed on the laser point cloud data corresponding to the multiple pillars.

[0092] Each point in each pillar can be represented by a 10-dimensional feature (x, y, z, r, xc, yc, zc, xp, yp, zp), where (xc, yc, zc) is the deviation of the current point from the average of all points in the current pillar, r is the intensity, and (xp, yp, zp) represent the distances of the current point from the center point of the current pillar in the (x, y, z) dimensions. Based on this, a certain strategy can be adopted to extract the features of the points in each pillar. For example, the dimension of the points can be increased to 64 dimensions using a fully connected layer, and then max pooling can be performed to compress the number of points in each pillar to 1, that is, only representative points are selected through max pooling.

[0093] Next, the data obtained above is mapped back to the original space and converted into a 2D pseudo-image, that is, the data is converted back to the original format. Specifically, the size (H, W) of the original space can be obtained first based on the size of each Pillar. Since the Pillar occupies the entire height, there is no height dimension here, that is, GridSize = [H, W, 1]. Then, based on the coordinates of each Pillar in the original data, the coordinates in the original space can be obtained, that is, the position of the current Pillar in the H*W space can be obtained.

[0094] In some embodiments of this application, the step of processing the laser point cloud data using a point cloud feature processing network in the target detection model to obtain processed point cloud features includes: dividing the point cloud space corresponding to the laser point cloud data to obtain multiple columns of the same size; extracting point cloud features from each column using multiple point cloud feature extraction strategies to obtain multiple point cloud features corresponding to each column; and concatenating the multiple point cloud features corresponding to each column to obtain the processed point cloud features corresponding to each column.

[0095] Because the original point cloud feature processing network in the PointPillars network mainly extracts features from each pillar by using fully connected layers and max pooling to obtain feature data for a single angle, some information loss still occurs. Therefore, this application improves the original point cloud feature processing network in the PointPillars network by adding multiple point cloud feature extraction strategies. This allows for the extraction of features from the pillar from different angles, increasing the semantic richness of the feature vectors and further improving the object detection performance.

[0096] In some embodiments of this application, the plurality of point cloud features includes a first point cloud feature, a second point cloud feature, and a third point cloud feature. The step of extracting point cloud features from each column using multiple point cloud feature extraction strategies to obtain the plurality of point cloud features corresponding to each column includes: using a fully connected layer in the point cloud feature processing network to increase the dimensionality of the points in each column, and using a max-pooling layer in the point cloud feature processing network to compress the number of points in each column, to obtain the first point cloud feature corresponding to each column; concatenating the points in each column into a matrix, and using a first convolutional layer in the point cloud feature processing network to convolve the matrix corresponding to each column, to obtain the second point cloud feature corresponding to each column; and using a second convolutional layer in the point cloud feature processing network to convolve the multidimensional features corresponding to each point in each column, to obtain the third point cloud feature corresponding to each column.

[0097] The improved point cloud feature processing network in this application mainly extracts point cloud features from three perspectives. The first perspective is based on the extraction strategy adopted by the original point cloud feature processing network. For example, the dimension of the points can be increased to 64 dimensions using a fully connected layer, and then after max pooling, the number of points in each pillar is compressed to 1. Representative points are selected to obtain the first point cloud features corresponding to each pillar.

[0098] The second approach involves adding an extraction strategy to the original point cloud feature processing network. This involves concatenating the 10-dimensional feature vectors of all points in each pillar into a matrix, resulting in a 32*10 matrix. Then, a convolutional kernel of the same size is used to convolve this matrix, yielding a 1*1 feature map with 64 channels, which serves as the second point cloud feature for each pillar. This extraction strategy essentially fuses the information of all points within each pillar, extracting more geometric information and thus improving subsequent object detection performance.

[0099] The third approach also involves adding an extraction strategy to the original point cloud feature processing network. This involves using a convolutional kernel of a certain size, such as a 1x3 kernel, to convolve the 10-dimensional features of each point in each pillar, obtaining the convolutional feature vector for each point, and finally obtaining the third point cloud features corresponding to each pillar. This strategy essentially fuses the multi-dimensional features corresponding to each point, thereby extracting richer semantic information.

[0100] Finally, the first, second, and third point cloud features of each Pillar extracted by the above three extraction strategies are stitched together, and the stitched data is then mapped back to the original space to be converted into a 2D pseudo-image, that is, the data is converted back into the original format.

[0101] In some embodiments of this application, the original feature pyramid network includes low-level convolutional layers and corresponding deconvolutional layers, as well as high-level convolutional layers and corresponding deconvolutional layers. The step of using the point cloud feature extraction network in the object detection model to extract features from the processed point cloud features to obtain the point cloud feature extraction result includes: adding a preset convolutional layer between the low-level convolutional layers and the corresponding deconvolutional layers; performing convolution processing on the processed point cloud features using the low-level convolutional layers to obtain a first convolution result of the low-level convolutional layers; performing convolution processing on the first convolution result of the low-level convolutional layers using the preset convolutional layers to obtain a second convolution result of the low-level convolutional layers; performing convolution processing on the first convolution result of the low-level convolutional layers using the high-level convolutional layers to obtain a convolution result of the high-level convolutional layers; and concatenating the second convolution result of the low-level convolutional layers with the convolution result of the high-level convolutional layers to obtain the point cloud feature extraction result.

[0102] like Figure 2 As shown, a schematic diagram of the original Feature Pyramid Network (FPN) structure in the prior art is provided. The original FPN network structure mainly includes low-level convolutional layers and corresponding deconvolutional layers, as well as high-level convolutional layers and corresponding deconvolutional layers. The number of low-level convolutional layers is generally two. Since the information of small targets is mainly stored in the low-level space, but the original FPN network has relatively few low-level convolutional layers, the resulting feature map contains a lot of noise information. Therefore, the embodiments of this application improve the network structure of the original Feature Pyramid Network, such as... Figure 3 As shown, a schematic diagram of an improved feature pyramid network in this application embodiment is provided. By adding a preset convolutional layer between the low-level convolutional layer and the corresponding deconvolutional layer, the extraction of semantic information in the low-level space is enhanced, noise is reduced, and the detection effect of small targets is improved.

[0103] In practice, the processed point cloud features can be convolved using low-scale convolutional layers to obtain the first convolution result. Then, on one hand, the first convolution result is used as input to a pre-defined convolutional layer to obtain the second convolution result, thus improving the extraction of small target information by adding convolutional processing to the low-scale space. On the other hand, this result is used as input to a high-scale convolutional layer to obtain the convolution result. Finally, the second convolution result of the low-scale convolutional layer and the convolution result of the high-scale convolutional layer are concatenated to obtain the final point cloud feature extraction result.

[0104] In some embodiments of this application, the original feature pyramid network includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. The step of using the point cloud feature extraction network in the object detection model to extract features from the processed point cloud features to obtain the point cloud feature extraction result includes: adding a first max-pooling layer between the first and second convolutional layers, and adding a second max-pooling layer between the second and third convolutional layers; performing convolution processing on the processed point cloud features using the first convolutional layer to obtain the convolution result of the first convolutional layer; and using the max-pooling layer between the first and second convolutional layers to perform convolution processing on the convolution result of the first convolutional layer. Pooling is performed to obtain the pooling result of the first max pooling layer; the pooling result of the first max pooling layer is convolved by the second convolutional layer to obtain the convolution result of the second convolutional layer; the convolution result of the second convolutional layer is pooled by the max pooling layer between the second and third convolutional layers to obtain the pooling result of the second max pooling layer; the pooling result of the second max pooling layer is convolved by the third convolutional layer to obtain the convolution result of the third convolutional layer; the convolution results of the first, second, and third convolutional layers are deconvolved and then concatenated to obtain the point cloud feature extraction result.

[0105] In addition to improving the detection performance of the target detection model for small targets, the embodiments of this application can also improve the original FPN network in the PointPillars network to further improve the detection performance of the target detection model for large targets.

[0106] like Figure 2As shown, the original FPN network mainly includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. The first and second convolutional layers can be considered as low-level convolutional layers, and the third convolutional layer can be considered as a high-level convolutional layer. To improve the model's detection performance for large targets, this embodiment can achieve this by adding a max pooling layer to the original FPN network, such as... Figure 4 As shown, a schematic diagram of another improved feature pyramid network in this application embodiment is provided. That is, a first max pooling layer can be added between the first convolutional layer and the second convolutional layer, and a second max pooling layer can be added between the second convolutional layer and the third convolutional layer. The pooling process of the max pooling layer can further increase the receptive field of the information contained in the obtained feature map, thereby improving the detection effect of large targets.

[0107] In some embodiments of this application, the step of using the point cloud feature extraction network in the target detection model to extract features from the processed point cloud features to obtain the point cloud feature extraction result includes: adding a parallel convolutional layer with a 1*1 kernel and a batch normalization (BN) layer to each convolutional layer of the original FPN network; then fusing the output of the 1*1 convolutional layer and the output of the batch normalization layer to obtain the output of that convolutional layer. This method increases the feature information extraction capability by adding branches, achieving the effect of model fusion.

[0108] In some embodiments of this application, the detection head network includes multiple detection heads. The step of using the detection head network in the target detection model to perform target detection on the point cloud feature extraction results to obtain target detection results includes: using multiple detection heads to detect the point cloud feature extraction results respectively, obtaining detection results from multiple detection heads, and concatenating the detection results from multiple detection heads to obtain a concatenated detection result; using multiple prediction branches in the detection head network to predict the concatenated detection result respectively, obtaining multiple prediction results as the target detection result, the multiple prediction results including category prediction result, bounding box prediction result, orientation prediction result, and foreground / background prediction result; the step of determining the loss value of the target detection result using a preset loss function, and updating the parameters of the target detection model based on the loss value of the target detection result includes: determining a category prediction loss based on the category prediction result, determining a bounding box prediction loss based on the bounding box prediction result, determining an orientation prediction loss based on the orientation prediction result, and determining a foreground / background prediction loss based on the foreground / background prediction result; and updating the parameters of the target detection model based on the category prediction loss, the bounding box prediction loss, the orientation prediction loss, and the foreground / background prediction loss.

[0109] like Figure 5 The diagram illustrates the structure of the original detection head network in a PointPillars network, a technique employed in the prior art. The original detection head network in PointPillars uses three detection heads similar to an SSD (Small Surface Diode). Anchors of different sizes and aspect ratios are set according to the category. The matching degree between the anchor (predicted bounding box) and the ground truth (real bounding box) is calculated. If the match exceeds a threshold, it is considered a positive sample; otherwise, it is a negative sample. Anchor generation defaults to predicting three categories per region, with two anchors per category. The output of the prediction head contains three types of data:

[0110] 1) Class prediction: For example, is the target category a vehicle, pedestrian, or bicycle?

[0111] 2) Box prediction: Each box has 7 dimensions, including the center point (x, y, z) coordinates, length, width, height h, w, l, and rotation angle r. If it is a vehicle target, the rotation angle r can be understood as the steering wheel angle.

[0112] 3) Direction prediction: Since the rotation angle loss in the box prediction uses the sine function, it cannot distinguish between positive and negative orientation boxes. Therefore, softmax is used to predict discrete rotation angle (orientation) categories.

[0113] Based on the three-dimensional data output by the original detection head network, this embodiment of the application further adds a set of output features, i.e., a fourth dimension of data, to the output layer to distinguish whether each Pillar belongs to the foreground or the background. For example... Figure 6 The diagram illustrates the structure of an improved detection head network according to an embodiment of this application. For locations where the output is a background point, the number of anchors can be reduced by lowering the output category score, thereby improving the overall efficiency of model training.

[0114] Correspondingly, during the loss function calculation stage, it is also necessary to calculate the predicted loss of the foreground / background points and use it as part of the total loss value, serving as another dimension for model updates and improving the overall efficiency of model training.

[0115] In addition, it should be noted that the improvements to the point cloud feature processing network, point cloud feature extraction network, and detection head network in the foregoing embodiments of this application can be combined with each other according to actual conditions, thereby further improving the target detection effect.

[0116] This application embodiment also provides a training device 700 for an object detection model, such as... Figure 7The diagram shows a schematic of a training device for an object detection model according to an embodiment of this application. The training device 700 for the object detection model includes: a point cloud feature processing unit 710, a point cloud feature extraction unit 720, a first detection unit 730, and an update unit 740, wherein:

[0117] The point cloud feature processing unit 710 is used to acquire laser point cloud data for training and process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features.

[0118] The point cloud feature extraction unit 720 is used to extract features from the processed point cloud features using the point cloud feature extraction network in the target detection model to obtain point cloud feature extraction results. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network.

[0119] The first detection unit 730 is used to perform target detection on the point cloud feature extraction result using the detection head network in the target detection model, and obtain the target detection result.

[0120] The update unit 740 is used to determine the loss value of the target detection result using a preset loss function, and update the parameters of the target detection model according to the loss value of the target detection result to obtain the trained target detection model.

[0121] In some embodiments of this application, the point cloud feature processing unit 710 is specifically used to: divide the point cloud space corresponding to the laser point cloud data to obtain multiple columns of the same size; use the farthest point sampling algorithm to sample the points in each column to obtain multiple columns after sampling processing; and perform point cloud feature processing on the laser point cloud data corresponding to the multiple columns after sampling processing to obtain the processed point cloud features.

[0122] In some embodiments of this application, the point cloud feature processing unit 710 is specifically used to: divide the point cloud space corresponding to the laser point cloud data to obtain multiple columns of the same size; extract point cloud features from each column using multiple point cloud feature extraction strategies to obtain multiple point cloud features corresponding to each column; and stitch together the multiple point cloud features corresponding to each column to obtain the processed point cloud features corresponding to each column.

[0123] In some embodiments of this application, the plurality of point cloud features include a first point cloud feature, a second point cloud feature, and a third point cloud feature. The point cloud feature processing unit 710 is specifically used to: increase the dimensionality of the points in each column using a fully connected layer in the point cloud feature processing network, and compress the number of points in each column using a max pooling layer in the point cloud feature processing network to obtain the first point cloud feature corresponding to each column; concatenate the points in each column into a matrix, and convolve the matrix corresponding to each column using a first convolutional layer in the point cloud feature processing network to obtain the second point cloud feature corresponding to each column; and convolve the multidimensional features corresponding to each point in each column using a second convolutional layer in the point cloud feature processing network to obtain the third point cloud feature corresponding to each column.

[0124] In some embodiments of this application, the original feature pyramid network includes low-level convolutional layers and corresponding deconvolutional layers, as well as high-level convolutional layers and corresponding deconvolutional layers. The point cloud feature extraction unit 720 is specifically used to: add a preset convolutional layer between the low-level convolutional layers and the corresponding deconvolutional layers; perform convolution processing on the processed point cloud features using the low-level convolutional layers to obtain a first convolution result of the low-level convolutional layers; perform convolution processing on the first convolution result of the low-level convolutional layers using the preset convolutional layers to obtain a second convolution result of the low-level convolutional layers; perform convolution processing on the first convolution result of the low-level convolutional layers using the high-level convolutional layers to obtain a convolution result of the high-level convolutional layers; and concatenate the second convolution result of the low-level convolutional layers and the convolution result of the high-level convolutional layers to obtain the point cloud feature extraction result.

[0125] In some embodiments of this application, the original feature pyramid network includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. The point cloud feature extraction unit 720 is specifically configured to: add a first max-pooling layer between the first and second convolutional layers, and add a second max-pooling layer between the second and third convolutional layers; perform convolution processing on the processed point cloud features using the first convolutional layer to obtain the convolution result of the first convolutional layer; and perform pooling processing on the convolution result of the first convolutional layer using the max-pooling layer between the first and second convolutional layers to obtain the pooling result of the first max-pooling layer. The following steps are performed: The pooling result of the first max pooling layer is convolved with the second convolutional layer to obtain the convolution result of the second convolutional layer; the convolution result of the second convolutional layer is pooled with the max pooling layer between the second and third convolutional layers to obtain the pooling result of the second max pooling layer; the pooling result of the second max pooling layer is convolved with the third convolutional layer to obtain the convolution result of the third convolutional layer; the convolution results of the first, second, and third convolutional layers are deconvolved and then concatenated to obtain the point cloud feature extraction result.

[0126] In some embodiments of this application, the detection head network includes multiple detection heads. The first detection unit 730 is specifically used to: use the multiple detection heads to detect the point cloud feature extraction results respectively, obtain the detection results of the multiple detection heads, and concatenate the detection results of the multiple detection heads to obtain the concatenated detection result; use multiple prediction branches in the detection head network to predict the concatenated detection result respectively, obtain multiple prediction results as the target detection result, the multiple prediction results including category prediction result, box prediction result, orientation prediction result, and foreground / background prediction result; the update unit 740 is specifically used to: determine the category prediction loss based on the category prediction result, determine the box prediction loss based on the box prediction result, determine the orientation prediction loss based on the orientation prediction result, and determine the foreground / background prediction loss based on the foreground / background prediction result; update the parameters of the target detection model based on the category prediction loss, the box prediction loss, the orientation prediction loss, and the foreground / background prediction loss.

[0127] It is understood that the training device for the above-mentioned object detection model can implement each step of the training method for the object detection model provided in the foregoing embodiments. The relevant explanations of the training method for the object detection model are applicable to the training device for the object detection model, and will not be repeated here.

[0128] This application also provides a target detection method, such as... Figure 8The diagram shows a flowchart of a target detection method according to an embodiment of this application. The target detection method includes at least the following steps S810 to S820:

[0129] Step S810: Obtain the current laser point cloud data;

[0130] Step S830: Use the target detection model to perform target detection on the current laser point cloud data to obtain the current target detection result;

[0131] The target detection model is trained based on the training method of any of the aforementioned target detection models.

[0132] In this embodiment of the application, when performing target detection, it is necessary to first acquire the currently collected laser point cloud data, and then input the currently collected laser point cloud data into the target detection model pre-trained in the aforementioned embodiment to obtain the target detection result.

[0133] This application embodiment also provides a target detection device 900, such as Figure 9 The diagram shows a schematic representation of a target detection device according to an embodiment of this application. The target detection device 900 includes: an acquisition unit 910 and a second detection unit 920, wherein:

[0134] Acquisition unit 910 is used to acquire the current laser point cloud data;

[0135] The second detection unit 920 is used to perform target detection on the current laser point cloud data using the target detection model to obtain the current target detection result;

[0136] The target detection model is trained based on the training device for the aforementioned target detection model.

[0137] It is understood that the target detection device described above can implement each step of the target detection method provided in the foregoing embodiments. The relevant explanations of the target detection method are applicable to the target detection device and will not be repeated here.

[0138] Figure 10 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Please refer to it. Figure 10 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0139] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus...

[0140] Lines, data buses, control buses, etc. For ease of representation, Figure 10 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0141] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0142] The processor reads the corresponding computer program from non-volatile memory into main memory and then runs it, forming a training device for the object detection model at the logical level 0. The processor executes the program stored in memory.

[0143] And specifically used to perform the following operations:

[0144] Acquire laser point cloud data for training, and process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features.

[0145] The point cloud feature extraction network in the target detection model is used to extract features from the processed point cloud feature 5 to obtain the point cloud feature extraction result. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network.

[0146] The target detection result is obtained by using the detection head network in the target detection model to perform target detection on the point cloud feature extraction result.

[0147] The loss value of the target detection result is determined by using a preset loss function, and the parameters of the target detection model are updated according to the loss value of the target detection result to obtain the trained target detection model.

[0148] The above is as stated in this application. Figure 1The method executed by the training device for the target detection model disclosed in the illustrated embodiment can be applied to a processor, or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method can be implemented by the processor.

[0149] The processor is implemented through integrated logic circuits in hardware or instructions in software. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0150] The electronic device can also perform Figure 1 The method for training a target detection model is described, and the training device for the target detection model is implemented in... Figure 1 The functions of the embodiments shown are not described in detail here.

[0151] This application also proposes a computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by an electronic device including multiple applications, enable the electronic device to perform... Figure 1 The method executed by the training device of the target detection model in the illustrated embodiment is specifically used to perform:

[0152] Acquire laser point cloud data for training, and process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features.

[0153] The point cloud feature extraction network in the target detection model is used to extract features from the processed point cloud features to obtain point cloud feature extraction results. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network.

[0154] The target detection result is obtained by using the detection head network in the target detection model to perform target detection on the point cloud feature extraction result.

[0155] The loss value of the target detection result is determined by using a preset loss function, and the parameters of the target detection model are updated according to the loss value of the target detection result to obtain the trained target detection model.

[0156] It should be noted that the electronic device in this application embodiment can also perform... Figure 8 The method executed by the target detection device, and the implementation of the target detection device in... Figure 8 The functions of the embodiments shown are not described in detail here.

[0157] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0158] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0159] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0160] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0161] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0162] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0163] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0164] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0165] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A training method for an object detection model, wherein, The training method for the object detection model includes: Acquire laser point cloud data for training, and process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features. The point cloud feature extraction network in the target detection model is used to extract features from the processed point cloud features to obtain point cloud feature extraction results. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network. The target detection result is obtained by using the detection head network in the target detection model to perform target detection on the point cloud feature extraction result. The loss value of the target detection result is determined by using a preset loss function, and the parameters of the target detection model are updated according to the loss value of the target detection result to obtain the trained target detection model; The original feature pyramid network includes low-level convolutional layers and corresponding deconvolutional layers, as well as high-level convolutional layers and corresponding deconvolutional layers. The point cloud feature extraction network in the target detection model is used to extract features from the processed point cloud features, resulting in the following point cloud feature extraction results: A preset convolutional layer is added between the low-level convolutional layer and the corresponding deconvolutional layer; The processed point cloud features are convolved using the low-level convolutional layer to obtain the first convolution result of the low-level convolutional layer. The first convolution result of the low-level convolutional layer is processed by the preset convolutional layer to obtain the second convolution result of the low-level convolutional layer. The first convolution result of the low-level convolutional layer is processed by the high-level convolutional layer to obtain the convolution result of the high-level convolutional layer. The point cloud feature extraction result is obtained by concatenating the second convolution result of the low-level convolutional layer and the convolution result of the high-level convolutional layer.

2. The training method for the target detection model as described in claim 1, wherein, The laser point cloud data is processed using a point cloud feature processing network in the target detection model to obtain processed point cloud features, including: The point cloud space corresponding to the laser point cloud data is divided to obtain multiple columnar bodies of the same size; The farthest point sampling algorithm is used to sample points in each column to obtain multiple columns after sampling. Point cloud feature processing is performed on the laser point cloud data corresponding to the multiple columnar bodies after sampling to obtain the processed point cloud features.

3. The training method for the target detection model as described in claim 1, wherein, The laser point cloud data is processed using a point cloud feature processing network in the target detection model to obtain processed point cloud features, including: The point cloud space corresponding to the laser point cloud data is divided to obtain multiple columnar bodies of the same size; Multiple point cloud feature extraction strategies are used to extract point cloud features from each column, resulting in multiple point cloud features corresponding to each column. The point cloud features corresponding to each column are stitched together to obtain the processed point cloud features corresponding to each column.

4. The training method for the target detection model as described in claim 3, wherein, The multiple point cloud features include a first point cloud feature, a second point cloud feature, and a third point cloud feature. The point cloud feature extraction is performed on each column using multiple point cloud feature extraction strategies to obtain multiple point cloud features corresponding to each column, including: The points in each column are increased in dimensionality using the fully connected layer in the point cloud feature processing network, and the number of points in each column is compressed using the max pooling layer in the point cloud feature processing network to obtain the first point cloud feature corresponding to each column. The points in each column are concatenated into a matrix, and the first convolutional layer in the point cloud feature processing network is used to convolve the matrix corresponding to each column to obtain the second point cloud feature corresponding to each column. The second convolutional layer in the point cloud feature processing network is used to convolve the multidimensional features corresponding to each point in each column to obtain the third point cloud features corresponding to each column.

5. The training method for the target detection model as described in claim 1, wherein, The original feature pyramid network includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. The point cloud feature extraction network in the target detection model is used to extract features from the processed point cloud features, resulting in the following point cloud feature extraction results: A first max pooling layer is added between the first convolutional layer and the second convolutional layer, and a second max pooling layer is added between the second convolutional layer and the third convolutional layer; The processed point cloud features are convolved using the first convolutional layer to obtain the convolution result of the first convolutional layer. The convolution result of the first convolutional layer is pooled using the max pooling layer between the first convolutional layer and the second convolutional layer to obtain the pooling result of the first max pooling layer. The pooling result of the first max pooling layer is convolved by the second convolutional layer to obtain the convolution result of the second convolutional layer; The convolution result of the second convolutional layer is pooled using the max pooling layer between the second and third convolutional layers to obtain the pooling result of the second max pooling layer. The pooling result of the second max pooling layer is convolved by the third convolutional layer to obtain the convolution result of the third convolutional layer; The convolution results of the first convolutional layer, the second convolutional layer, and the third convolutional layer are deconvolved and then concatenated to obtain the point cloud feature extraction result.

6. The training method for the target detection model as described in claim 1, wherein, The detection head network includes multiple detection heads. The target detection model utilizes the detection head network within the target detection model to perform target detection on the point cloud feature extraction results, obtaining target detection results including: The point cloud feature extraction results are detected by multiple detection heads respectively to obtain the detection results of multiple detection heads, and the detection results of multiple detection heads are stitched together to obtain the stitched detection result; The concatenated detection results are predicted by multiple prediction branches in the detection head network to obtain multiple prediction results, which are used as the target detection results. The multiple prediction results include category prediction results, bounding box prediction results, orientation prediction results, and foreground / background prediction results. The step of determining the loss value of the target detection result using a preset loss function and updating the parameters of the target detection model based on the loss value of the target detection result includes: The category prediction loss is determined based on the category prediction result, the box prediction loss is determined based on the box prediction result, the direction prediction loss is determined based on the direction prediction result, and the foreground / background prediction loss is determined based on the foreground / background prediction result. The parameters of the object detection model are updated based on the category prediction loss, the bounding box prediction loss, the orientation prediction loss, and the foreground / background prediction loss.

7. A target detection method, wherein, The target detection method includes: Obtain the current laser point cloud data; The target detection model is used to perform target detection on the current laser point cloud data to obtain the current target detection result; The target detection model is trained based on the training method of any one of the target detection models in claims 1 to 6.

8. A training device for an object detection model, wherein, The training device for the target detection model includes: The point cloud feature processing unit is used to acquire laser point cloud data for training and to process the laser point cloud data using the point cloud feature processing network in the target detection model to obtain processed point cloud features. The point cloud feature extraction unit is used to extract features from the processed point cloud features using the point cloud feature extraction network in the target detection model, and obtain the point cloud feature extraction result. The point cloud feature extraction network is obtained by adding convolutional layers to the low-level scale space of the original feature pyramid network. The first detection unit is used to perform target detection on the point cloud feature extraction results using the detection head network in the target detection model, and obtain the target detection result. The update unit is used to determine the loss value of the target detection result using a preset loss function, and update the parameters of the target detection model according to the loss value of the target detection result to obtain the trained target detection model. The original feature pyramid network includes low-level convolutional layers and corresponding deconvolutional layers, as well as high-level convolutional layers and corresponding deconvolutional layers. The point cloud feature extraction unit is specifically used for: A preset convolutional layer is added between the low-level convolutional layer and the corresponding deconvolutional layer; The processed point cloud features are convolved using the low-level convolutional layer to obtain the first convolution result of the low-level convolutional layer. The first convolution result of the low-level convolutional layer is processed by the preset convolutional layer to obtain the second convolution result of the low-level convolutional layer. The first convolution result of the low-level convolutional layer is processed by the high-level convolutional layer to obtain the convolution result of the high-level convolutional layer. The point cloud feature extraction result is obtained by concatenating the second convolution result of the low-level convolutional layer and the convolution result of the high-level convolutional layer.

9. A target detection device, wherein, The target detection device includes: The acquisition unit is used to acquire the current laser point cloud data; The second detection unit is used to perform target detection on the current laser point cloud data using the target detection model, and obtain the current target detection result. The target detection model is trained based on the training device of the target detection model of claim 8.

10. An electronic device, comprising: processor; as well as A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the training method of any one of the target detection models of claims 1 to 6, or to perform the target detection method of claim 7.

11. A computer-readable storage medium storing one or more programs, which, when executed by an electronic device including a plurality of applications, cause the electronic device to perform a training method for the object detection model of any one of claims 1 to 6, or to perform the object detection method of claim 7.

Images