A detection frame determination method, device, equipment and storage medium

Abstract

A detection box determination method, device, equipment and storage medium are disclosed. The method comprises: determining a first target point feature and a semantic feature; inputting the first target point feature into a first preset convolutional network to determine a confidence degree of a convolutional point cloud feature belonging to a foreground point feature; determining an initial key point, and determining a target key point feature based on the confidence degree, the initial key point and the distance of each point in the target point cloud except the initial key point; fusing the semantic feature, the target key point feature and the first target point feature to obtain a second target point feature; processing the second target point feature based on a second preset convolutional network to obtain a candidate detection box, and determining a target detection box according to the difference between the size of the candidate detection box and the size of a preset box. The technical scheme of the embodiment of the present application retains the three-dimensional spatial information of the point cloud, solves the problem of a small number of foreground points, enhances the quality of the foreground point cloud, and improves the accuracy of the detection box.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target object detection, and in particular to a bounding box determination method and device, equipment and a storage medium. BACKGROUND

[0002] With the continuous development of unmanned driving technology, a laser radar is generally installed on a vehicle capable of automatic driving, and through processing of point cloud data collected by the laser radar, detection and tracking of a three-dimensional target object can be realized.

[0003] At present, existing detection of a three-dimensional target object can be roughly divided into two categories, one is a voxel-based feature extraction algorithm, and the other is a raw point cloud-based feature extraction algorithm. Among them, the voxel-based feature extraction algorithm generally converts unstructured point cloud data into three-dimensional voxels or into a two-dimensional overhead view grid, and the raw point cloud-based feature extraction algorithm directly extracts features from raw point clouds by using a point cloud network, without voxelization and convolution downsampling of the point clouds, and high-precision three-dimensional spatial position information of the raw point clouds is retained.

[0004] However, in the downsampling process of the raw point cloud-based feature extraction algorithm, three-dimensional spatial information of the raw point clouds is lost, resulting in low accuracy of a generated bounding box, and the number of foreground points extracted by the raw point cloud-based feature extraction algorithm is small, resulting in low accuracy of the generated bounding box. SUMMARY

[0005] The present application provides a bounding box determination method, device, equipment and storage medium to solve the problem of low accuracy of a bounding box generated from point clouds.

[0006] In a first aspect, an embodiment of the present application provides a bounding box determination method, comprising:

[0007] determining a first target point feature and a semantic feature of a target point cloud;

[0008] inputting the first target point feature into a first preset convolutional network, and determining a confidence degree of each feature in a convolutional point cloud feature belonging to a foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network at least includes a sparse convolutional layer, a sparse deconvolutional layer and a manifold sparse convolutional layer;

[0009] determining an initial key point from the target point cloud, and determining a target key point feature based on the confidence degree, the initial key point and a distance of each point in the target point cloud except the initial key point;

[0010] fusing the semantic feature, the target key point feature and the first target point feature to obtain a second target point feature;

[0011] The second target point feature is a three-dimensional feature.

[0012] In a second aspect, an embodiment of the present application provides a detection frame determination device, comprising:

[0013] A first feature determination module is configured to determine a first target point feature and a semantic feature of a target point cloud.

[0014] A confidence determination module is configured to input the first target point feature into a first preset convolutional network, and determine a confidence of each feature in a convolutional point cloud feature belonging to a foreground point feature according to an output of the first preset convolutional network.

[0015] A second feature determination module is configured to determine an initial key point from the target point cloud, and determine a target key point feature based on the confidence, the initial key point, and a distance between each point in the target point cloud except the initial key point.

[0016] A third feature determination module is configured to fuse the semantic feature, the target key point feature, and the first target point feature to obtain a second target point feature.

[0017] A detection frame determination module is configured to process the second target point feature based on a second preset convolutional network to obtain a candidate detection frame, and determine whether to determine the candidate detection frame as a target detection frame according to a difference between a size of the candidate detection frame and a size of a preset frame, wherein the second target point feature is a three-dimensional feature.

[0018] In a third aspect, an embodiment of the present application provides an electronic device, comprising:

[0019] at least one processor;

[0020] and a memory in communication with the at least one processor;

[0021] wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the detection frame determination method of the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium storing computer instructions for causing a processor to implement the method for determining a detection box of the first aspect.

[0023] The method for determining a detection box provided by the embodiment of the present application determines a first target point feature and a semantic feature of a target point cloud, inputs the first target point feature into a first preset convolutional network, determines a confidence that each feature in a convolutional point cloud feature belongs to a foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network at least includes a sparse convolutional layer, a sparse deconvolutional layer and a manifold sparse convolutional layer, determines an initial key point from the target point cloud, and determines a target key point feature based on the confidence, the initial key point and a distance of each point in the target point cloud except the initial key point, fuses the semantic feature, the target key point feature and the first target point feature to obtain a second target point feature, processes the second target point feature based on a second preset convolutional network to obtain a candidate detection box, and determines whether to determine the candidate detection box as a target detection box according to a difference between a size of the candidate detection box and a size of a preset box, wherein the second target point feature is a three-dimensional feature. By using the above technical solution, the first target point feature of the target point cloud is input into the first preset convolutional network to obtain the confidence that each feature in the convolutional point cloud feature belongs to the foreground point feature, then the target key point feature is determined according to the confidence and the distance between the initial key point in the target point cloud and other points in the target point cloud, then the semantic feature of the target point cloud, the target key point feature and the first target point feature are fused to obtain the second target point feature, finally the second preset convolutional network is used to process the second target point feature to obtain the candidate detection box, and the target detection box is determined according to the difference between the size of the candidate detection box and the size of the preset box. Compared with the traditional method for determining a detection box, the three-dimensional spatial information of the point cloud is retained, the confidence determined by the preset convolutional network can solve the problem of a small number of foreground points in the target key point, the quality of the foreground point cloud is enhanced, sufficient feature information is provided for subsequent generation of a detection box, and the accuracy of the detection box is improved.

[0024] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present application, nor to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to make the technical solutions in the embodiments of the present application clearer, the accompanying drawings needed in the embodiment description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other embodiments obtained by those of ordinary skill in the art without any creative work based on the accompanying drawings should be within the protection scope of the present application.

[0026] Figure 1 is a flow chart of a detection frame determination method according to an embodiment of the present application;

[0027] Figure 2 is a flow chart of a detection frame determination method according to an embodiment of the present application;

[0028] Figure 3 is a training framework diagram according to an embodiment of the present application;

[0029] Figure 4 is a training process visualization diagram in a KITTI dataset according to an embodiment of the present application;

[0030] Figure 5 is a structural schematic diagram of a detection frame determination device according to an embodiment of the present application;

[0031] Figure 6 is a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the technical solutions in the embodiments of the present application clearer, the accompanying drawings needed in the embodiment description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other embodiments obtained by those of ordinary skill in the art without any creative work based on the accompanying drawings should be within the protection scope of the present application.

[0033] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In the description of the present application, "a plurality of" means two or more, unless otherwise specified. The association relationship of the associated objects is described, and it means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. The character " / " generally represents an "or" relationship between the front and rear associated objects. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0034] Embodiment one

[0035] Figure 1 A flowchart of a method for determining a detection frame is provided for the first embodiment of the present application. The present embodiment can be applicable to the case of generating a detection frame based on point cloud data. The method can be executed by a detection frame determination device, which can be realized in the form of hardware and / or software. The detection frame determination device can be configured in an electronic device, which can be composed of two or more physical entities or one physical entity.

[0036] As Figure 1 shown, the method for determining a detection frame provided by the first embodiment of the present application specifically includes the following steps:

[0037] S101, determining a first target point feature and a semantic feature of a target point cloud.

[0038] In the present embodiment, the point cloud data can be acquired by a radar device such as a laser radar, which is the initial point cloud. Then, the initial point cloud can be processed by a first preset method, such as selecting points in the initial point cloud as target points according to a certain proportion, so as to obtain a target point cloud. The three-dimensional coordinates of each point in the target point cloud are the target point features, i.e. the first target point features, and the semantic features of the target point cloud are extracted. Among them, since there are multiple ways to extract semantic features, this is not limited here. The point cloud can be understood as a collection of point data, and the points in the point cloud are usually represented in the form of three-dimensional coordinates, i.e. the point cloud contains multiple three-dimensional coordinates, which are point features.

[0039] S102, input the first target point feature into a first preset convolutional network, and determine the confidence that each feature in the convolutional point cloud feature belongs to the foreground point feature according to the output of the first preset convolutional network, wherein the first preset convolutional network at least includes a sparse convolutional layer, a sparse deconvolutional layer, and a popular sparse convolutional layer manifold sparse convolutional layer.

[0040] In the embodiment, the first preset convolutional network can be constructed to calculate the confidence, and the first preset convolutional network can include a sparse convolutional layer, a sparse deconvolutional layer, and a popular sparse convolutional layer manifold sparse convolutional layer. After the first target point feature is processed by the sparse convolutional layer, fine-grained feature information in the first target point feature is extracted. Then, the fine-grained feature information is processed by the sparse deconvolutional layer, and the fine-grained feature information is restored to initial feature information, i.e., to the first target point feature. The feature information after the deconvolutional processing is input into the popular sparse convolutional layer manifold sparse convolutional layer to obtain the convolutional point cloud feature and the confidence that the convolutional point cloud feature belongs to the foreground point feature. The convolutional point cloud feature and the first target point feature can usually correspond one-to-one. If there is a case of not corresponding, it can be regarded as error data and ignored.

[0041] S103, determining an initial key point from the target point cloud, and determining a target key point feature based on the confidence, the initial key point, and the distance between each point in the target point cloud except the initial key point.

[0042] In the embodiment, the initial key point can be determined from the target point cloud by using a second preset method, such as selecting at least two points from the target point cloud as the initial key point. The distance between the initial key point and each point in the target point cloud except the initial key point is calculated. Then, the points that meet the preset requirements are determined from the distances. The points that meet the preset requirements are the target key points, and the features of the points are the target key point features. The preset requirements include the requirement for the confidence and the requirement for the distance, such as the confidence exceeding a preset confidence value and the distance exceeding a preset distance value.

[0043] S104, fusing the semantic feature, the target key point feature, and the first target point feature to obtain a second target point feature.

[0044] In the embodiment, the fusion method can be splicing, i.e., the semantic feature, the target key point feature, and the first target point feature can be spliced to obtain the second target point feature.

[0045] S105, processing the second target point feature based on a second preset convolutional network to obtain a candidate detection frame, and determining whether to determine the candidate detection frame as a target detection frame according to a difference between a size of the candidate detection frame and a size of a preset frame, wherein the second target point feature is a three-dimensional feature.

[0046] In the embodiment, the second preset convolutional network can be used to extract a plurality of key features in the second target point feature, and a candidate frame can be generated according to the similarity and / or correlation degree of the plurality of key features, for example, key features with high similarity are determined as features in the same candidate frame, and then the difference between the size of the candidate detection frame and the size of the preset frame is compared. If the difference meets the set requirement, for example, the difference between the length, width and height of the candidate detection frame and the length, width and height of the preset frame is less than a preset difference value, the candidate detection frame can be determined as the target detection frame. The detection frame is generally used to frame the object to be detected, such as a person, a vehicle and a plant, so as to perform object detection and / or object tracking on the object in the detection frame. The type of the preset frame can be various, such as a vehicle frame and a person frame. The size of the preset frame can be set according to the actual situation, such as setting the size of the vehicle frame according to the size of the conventional vehicle.

[0047] The method for determining a detection frame provided by the embodiment of the present application determines a first target point feature and a semantic feature of a target point cloud, inputs the first target point feature into a first preset convolutional network, determines a confidence degree that each feature in a convolutional point cloud feature belongs to a foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network at least includes a sparse convolutional layer, a sparse deconvolutional layer, and a popular sparse convolutional layer manifold sparse convolutional layer, determines an initial key point from the target point cloud, and determines a target key point feature based on the confidence degree, the initial key point, and a distance of each point in the target point cloud except the initial key point, fuses the semantic feature, the target key point feature, and the first target point feature to obtain a second target point feature, processes the second target point feature based on a second preset convolutional network to obtain a candidate detection frame, and determines whether to determine the candidate detection frame as a target detection frame according to a difference between a size of the candidate detection frame and a size of a preset frame, wherein the second target point feature is a three-dimensional feature. The technical scheme of the embodiment of the present application inputs the first target point feature of the target point cloud into the first preset convolutional network to obtain the confidence degree that each feature in the convolutional point cloud feature belongs to the foreground point feature, then determines the target key point feature according to the confidence degree and the distance between the initial key point and other points in the target point cloud except the initial key point, fuses the semantic feature of the target point cloud, the target key point feature, and the first target point feature to obtain the second target point feature, finally processes the second target point feature based on the second preset convolutional network to obtain the candidate detection frame, and determines the target detection frame according to the difference between the size of the candidate detection frame and the size of the preset frame. Compared with the traditional method for determining a detection frame, the three-dimensional spatial information of the point cloud is retained, the confidence degree determined by the preset convolutional network can solve the problem that the number of foreground points in the target key point is small, the quality of the foreground point cloud is enhanced, sufficient feature information is provided for subsequent generation of a detection frame, and the accuracy of the detection frame is improved.

[0048] Embodiment Two

[0049] Figure 2 The flowchart of the method for determining a detection frame provided by the embodiment Two of the present application, the technical scheme of the embodiment of the present application is further optimized on the basis of the above-mentioned optional technical schemes, and a specific way of generating a detection frame is given.

[0050] Optionally, the first target point feature is input into a first preset convolutional network, and a confidence that each feature in the convolutional point cloud feature belongs to the foreground point feature is determined according to an output of the first preset convolutional network, including: performing down-sampling processing on the first target point feature by using a sparse convolutional layer of the first preset convolutional network to obtain a first feature; performing up-sampling processing on the first feature by using a sparse deconvolutional layer of the first preset convolutional network to obtain a second feature; inputting the second feature into a popular sparse convolutional layer of the first preset convolutional network, and determining the confidence that each feature in the convolutional point cloud feature belongs to the foreground point feature according to an output of a nonlinear classifier in the popular sparse convolutional layer. The advantage of this setting is that the feature data obtained after up-sampling and down-sampling contains more fine-grained features, making the calculation of the confidence more accurate, thereby improving the accuracy of the detection frame generated subsequently.

[0051] Optionally, the initial key points are determined from the target point cloud, and a target key point feature is determined based on the confidence, and distances between the initial key points and each point in the target point cloud except the initial key points, including: determining a first initial key point and a second initial key point from the target point cloud; determining a minimum value of a Euclidean distance between each point in the target point cloud except the initial key points and the initial key points as a first Euclidean distance to obtain a first Euclidean distance set, wherein the initial key points include the first initial key point and the second initial key point; determining a weight coefficient of each first Euclidean distance in the first Euclidean distance set according to a size of the confidence, and calculating a product of the first Euclidean distance and the weight coefficient to obtain a second Euclidean distance set, wherein the weight coefficient is positively correlated with the confidence; before a number of the target key point features reaches a preset number, determining a first target point feature corresponding to a maximum second Euclidean distance in the second Euclidean distance set as a newly added target key point feature; wherein an initial value of the number of the target key point features is zero, and the weight coefficient is adjusted once for each newly added target key point feature, and the second Euclidean distance set is updated according to the adjustment of the weight coefficient. The advantage of this setting is that the weight coefficient of the first Euclidean distance is adjusted according to the size of the confidence, and the Euclidean distance of the foreground point will be greater than that of the background point, so that more foreground points can be determined as target key points, that is, more foreground points are extracted, and the problem of inaccurate foreground point extraction is solved.

[0052] Optionally, the fusing the semantic feature, the target key point feature and the first target point feature to obtain a second target point feature comprises: splicing the target key point feature and the first target point feature to obtain a third feature; and using a third preset convolutional network to down-sample the third feature into a fourth feature matching the dimension of the semantic feature, and splicing the fourth feature and the semantic feature to obtain the second target point feature. The advantage of this setting is that the target key point feature and the first target point feature are spliced first, and then the spliced result is down-sampled to match the dimension of the semantic feature, and then spliced, so that the three-dimensional space feature and the semantic feature are better fused, and the fusion result, i.e. the second target point feature, contains more complete three-dimensional space features.

[0053] As shown in Figure 2 The method for determining a detection frame provided by the second embodiment of the present application specifically comprises the following steps:

[0054] S201, determining a first target point feature and a semantic feature of a target point cloud.

[0055] Optionally, the determination manner of the first target point feature comprises: voxelizing an initial point cloud in a preset radar coordinate system to obtain a voxel grid, wherein the voxel grid comprises a plurality of voxel small grids; calculating the mean value of the coordinates of the points in the non-empty voxel small grids, determining the mean value as a target point coordinate, and generating a plurality of target points according to the target point coordinate, wherein the voxel small grids comprise empty voxel small grids and non-empty voxel small grids, and each non-empty voxel small grid comprises at least one point in the initial point cloud; determining a target point cloud according to the plurality of target points, and determining the target point coordinate as the first target point feature of the target point cloud. The advantage of this setting is that the target point cloud is approximately equivalent to the initial point cloud, thereby reducing the workload of subsequent convolution processing.

[0056] Specifically, the initial point cloud can be divided into a plurality of voxel small grids in units of voxels, thereby obtaining a voxel grid, each voxel small grid being a voxel, non-empty voxel small grids are selected from the plurality of voxel small grids, and the mean value of the coordinates of the points contained in each non-empty voxel small grid is calculated. For example, if two points in a non-empty voxel small grid have coordinates (1, 2, 3) and (4, 5, 6), the mean value is (2.5, 3.5, 4.5), which is the target point coordinate and the first target point feature. The point corresponding to the target point coordinate is the target point, and a plurality of target points can form a target point cloud. The voxel is a volume element, which can be understood as a three-dimensional pixel. Each non-empty voxel small grid can contain a plurality of points in the initial point cloud, i.e. a plurality of three-dimensional point coordinates. The resolution of the voxel grid can be preset, such as 70x30x2.

[0057] S202, performing down-sampling processing on the first target point feature by using a sparse convolution layer of the first preset convolution network to obtain a first feature.

[0058] Specifically, the first target point feature can be input into the sparse convolution layer of the first preset convolution network to obtain the first feature. The process of the sparse convolution layer processing the first target point feature is a down-sampling process. The related parameters of the sparse convolution layer can be preset, such as setting the convolution kernel size of the sparse convolution layer to 3x3x3, the stride to 2, the down-sampling factor to 8, and the dimension of the output first feature to 16, 32, 64, and 64.

[0059] S203, performing up-sampling processing on the first feature by using a sparse deconvolution layer of the first preset convolution network to obtain a second feature.

[0060] Specifically, the first feature can be input into the sparse deconvolution layer of the first preset convolution network to obtain the second feature. The process of the sparse deconvolution layer processing the first feature is an up-sampling process. The processing of the sparse deconvolution layer can be approximately understood as the inverse processing of the sparse convolution. The second feature has a corresponding relationship with the first target point feature, but the second feature contains more fine-grained features than the first target point feature.

[0061] S204, inputting the second feature into a popular sparse convolution layer manifold sparse convolution layer of the first preset convolution network, and determining the confidence that each feature in the convolution point cloud feature belongs to the foreground point feature according to the output of a nonlinear classifier in the popular sparse convolution layer manifold sparse convolution layer.

[0062] Specifically, the second feature can be input into the popular sparse convolution layer manifold sparse convolution layer of the first preset convolution network. After processing by the nonlinear classifier in the popular sparse convolution layer manifold sparse convolution layer, the convolution point cloud feature, the convolution point cloud, and the confidence that each convolution point cloud feature belongs to the foreground point can be obtained. Since each point in the convolution point cloud has a corresponding relationship with each point in the target point cloud, the confidence of the convolution point cloud feature also has a corresponding relationship with each point in the target point cloud. The type of the nonlinear classifier is not limited, such as a sigmoid function. The output confidence is usually a value between 0 and 1. The greater the confidence, the greater the probability that the convolution point cloud feature belongs to the foreground point.

[0063] Optionally, in the pre-training stage, the loss function of the nonlinear classifier of the first preset convolution network can be which can be a focal loss function, wherein:

[0064]

[0065] s represents the confidence that the ith feature belongs to the foreground point feature, a = 0.15, g = 2, N represents the number of points in the convolution point cloud, and n represents the number of features with a confidence greater than a preset confidence threshold, such as the number of features with a confidence greater than 0.5.

[0066] S205, determining a first initial key point and a second initial key point from the target point cloud.

[0067] Specifically, two points in the target point cloud that satisfy a preset key point requirement, such as a distance of more than 2 meters, can be determined as initial key points, i.e., a first initial key point and a second initial key point.

[0068] Optionally, the determining of the first initial key point and the second initial key point from the target point cloud includes: determining a first initial key point from the target point cloud, and determining a first target point feature corresponding to the first initial key point in the target point cloud as a first initial key point feature; determining a third Euclidean distance between the first initial key point and each point in the target point cloud except the first initial key point according to the first initial key point feature and a first target point feature corresponding to each point in the target point cloud except the first initial key point, to obtain a third Euclidean distance set; determining a first target point feature corresponding to a maximum third Euclidean distance in the third Euclidean distance set as a second initial key point feature, and determining a second initial key point according to the second initial key point feature. The advantage of this setting is that compared with determining only one key point, first determining a key point and then determining another key point with the farthest Euclidean distance from the key point improves the probability of selection of foreground points.

[0069] Specifically, a random point in the target point cloud can be selected as a first initial key point, and a feature of the point in the target point cloud, i.e., a first target point feature, is determined as a first initial key point feature. Then, a Euclidean distance between the first initial key point and other points in the target point cloud is calculated according to the first initial key point feature and the first target point feature, i.e., a third Euclidean distance between the first initial key point and each point in the target point cloud except the first initial key point. A corresponding Euclidean distance set can be generated according to the determined multiple Euclidean distances. The maximum Euclidean distance is filtered out from the Euclidean distance set. The first target point feature corresponding to the maximum Euclidean distance is the second initial key point feature, and the point corresponding to the second initial key point feature is the second initial key point. The first target point feature and the first initial key point feature are both three-dimensional point coordinates, so the Euclidean distance between the first initial key point and the points in the target point cloud can be obtained by calculation.

[0070] Optionally, the first initial key point can be selected in the target point cloud, and the second initial key point can be determined in a circle with the first initial key point as the center and a preset value as the radius.

[0071] In S206, a minimum value of the Euclidean distance between each point in the target point cloud except the initial key point and the initial key point is determined as a first Euclidean distance, and a first Euclidean distance set is obtained.

[0072] The initial key point includes the first initial key point and the second initial key point.

[0073] Specifically, the minimum Euclidean distance between each point in the target point cloud except the initial key point and the two initial key points, i.e., the first initial key point and the second initial key point, can be calculated, and thus a first Euclidean distance set composed of the minimum Euclidean distance is obtained.

[0074] In S207, a weight coefficient of each first Euclidean distance in the first Euclidean distance set is determined according to the size of the confidence, and a product of the first Euclidean distance and the weight coefficient is calculated, and thus a second Euclidean distance set is obtained.

[0075] The weight coefficient is positively correlated with the confidence.

[0076] Specifically, the weight coefficient can be set according to the size of the confidence corresponding to the point in the target point cloud, and the greater the confidence, the greater the weight coefficient. Then, the product of each first Euclidean distance in the first Euclidean distance set and the weight coefficient is calculated, i.e., the first Euclidean distance is weighted and calculated, and thus a second Euclidean distance set composed of the product is obtained. The confidence of the convolution point cloud feature has a corresponding relationship with each point in the target point cloud, and the effect of the weighted calculation of the first Euclidean distance can be understood as relatively increasing the Euclidean distance of the foreground point and relatively reducing the Euclidean distance of the background point.

[0077] In S208, when the number of the determined target key point features does not reach a preset number, a first target point feature corresponding to the maximum second Euclidean distance in the second Euclidean distance set is determined as a newly added target key point feature.

[0078] The initial value of the number of the target key point features is zero, the weight coefficient is adjusted once each time a newly added target key point feature is determined, and the second Euclidean distance set is updated according to the adjustment of the weight coefficient.

[0079] Specifically, the number of target key point features can be set, i.e., a preset number, such as 5. When the number of target key point features is less than 5, the first target point feature corresponding to the second Euclidean distance with the largest second Euclidean distance in the second Euclidean distance set can be determined as a target key point feature. When the number of target key point features is equal to 5, no new target key point feature is determined. After each new target key point feature is determined, the weight coefficient can be adjusted once, and the product of the weight coefficient and the first Euclidean distance is updated synchronously, i.e., the second Euclidean distance set is updated synchronously. The adjusted weight coefficient still follows the principle that the greater the confidence, the greater the weight coefficient.

[0080] S209, splicing the target key point features and the first target point features to obtain third features.

[0081] Specifically, the target key point features and the first target point features can be spliced first to obtain third features.

[0082] S210, using a third preset convolutional network to down-sample the third features into fourth features matching the dimension of the semantic features, and splicing the fourth features and the semantic features to obtain second target point features.

[0083] Specifically, after the third features are obtained, the third features can be down-sampled using a third preset convolutional network to obtain fourth features, and then the fourth features and the semantic features of the target point cloud are spliced to obtain second target point features. The third preset convolutional network includes a three-dimensional sparse convolutional layer.

[0084] S211, processing the second target point features based on a second preset convolutional network to obtain a candidate bounding box, and determining whether the candidate bounding box is determined as a target bounding box according to the difference between the size of the candidate bounding box and the size of a preset bounding box.

[0085] Optionally, the processing of the second target point features based on the second preset convolutional network to obtain a candidate bounding box, and determining whether the candidate bounding box is determined as a target bounding box according to the difference between the size of the candidate bounding box and the size of a preset bounding box includes: compressing the second target point features to obtain a two-dimensional top view compressed feature; clustering the two-dimensional top view compressed feature using the second preset convolutional network to obtain a candidate bounding box; and determining whether the intersection over union of the size of the candidate bounding box and the size of a preset bounding box is greater than a set threshold value. If yes, the candidate bounding box is determined as a target bounding box. This setting has the advantage that the difference between the size of the candidate bounding box and the size of a preset bounding box can be accurately determined by calculating the intersection over union, so that the generated target bounding box is more accurate.

[0086] Specifically, the three-dimensional second target point feature can be compressed into a two-dimensional top view feature, i.e., a two-dimensional top view compressed feature, and then a second preset convolutional network can be used to extract fine-grained features in the two-dimensional top view compressed feature, and the fine-grained features can be clustered to generate a candidate detection box. Finally, the intersection over union of the size of the candidate detection box and the size of the preset box is calculated. If the intersection over union is greater than a set threshold, it indicates that the size of the candidate detection box is close to the size of the preset box, and the candidate detection box can be determined as the target detection box. After generating the candidate detection box, the type of the candidate detection box can be identified using a preset model or algorithm, and then a preset box consistent with the type of the candidate detection box can be determined. The type of the preset box usually includes multiple types, and the size of the preset box corresponding to different types is different. For example, the size of the car box is 3.9 meters long, 1.6 meters wide, and 1.56 meters high, and the size of the pedestrian box is 0.8 meters long, 0.6 meters wide, and 1.7 meters high. The set threshold for different types of preset boxes can be different. For example, the set threshold for the car box can be 0.6, and the set threshold for the pedestrian box can be 0.5.

[0087] Optionally, Figure 3 A training framework, in the early training stage, the determination model of the target detection box can be trained based on the self-supervised learning mechanism to make the accuracy of the target detection box accurate enough. The process of generating the target detection box by the determination model of the target detection box corresponds to the method of determining the detection box in this paper. The training process is shown in Figure 3 When training the determination model of the target detection box, a standard box, also known as an anchor box, needs to be labeled in the sample point cloud by artificial marking first, and then the labeled point cloud is input into the point cloud segmentation module, i.e., the first preset convolutional network. The confidence of the output is determined as the key point feature, i.e., the target key point feature. The semantic feature, key point feature, and spatial feature, i.e., the first target point feature, are spliced to obtain the spliced feature, i.e., the second target point feature. The spliced feature is input into the fusion module, i.e., the second target point feature is regressed according to the labeled box to obtain the candidate box, and the type of the candidate box is classified. Whether the candidate box is determined as the target detection box is determined by the intersection over union of the size of the candidate box of the same type and the size of the anchor box. The target detection box generated during training can be regarded as a real box, the candidate detection box generated during training can be regarded as a candidate box, and the region generation network loss function and the regression loss function are calculated.

[0088]

[0089] represents a class loss value, i.e., a class loss value of a candidate bounding box, which is determined according to a class loss function, the class loss function can be a Focal loss function, and β is a coefficient, represents a residual between the candidate box and the anchor box, represents a residual between the real box and the anchor box, represents a loss value between the candidate box and the real box, which can be determined by cross-entropy calculation, the more similar the candidate box and the real box are, the smaller, represents a residual between the candidate box and the real box, represents a residual between the candidate box and the anchor box, the candidate box, the real box and the anchor box can be represented in the form of {x, y, z, l, h, w, θ}, (x, y, z) represents the center coordinates of the candidate box, the real box and the anchor box in the radar coordinate system of the two-dimensional top view, l, h and w represent the length, width and height dimensions of the candidate box, the real box and the anchor box, and θ represents the angle between the object forward direction of the candidate box and the horizontal axis in the camera coordinate system. The total loss function L is, When the value of the total loss function meets the set loss requirement, such as being less than 0.2, it means that the training is completed, i.e., the accuracy of the real box is accurate enough. Wherein, Figure 4 is a training process visualization diagram in a KITTI dataset, and the related sample data and training data during the training can be stored in a preset storage area, such as in the KITTI dataset, the training process visualization is as shown in Figure 4 .

[0090] The determination method of the detection box provided by the embodiment of the application determines the confidence that the convolution point cloud feature belongs to the foreground point feature by using the first preset convolution network, then determines the initial key point in the target point cloud, sets the weight coefficient of the first target point feature according to the size of the confidence, determines the target key point feature according to the product of the weight coefficient and the distance between the initial key point and other points in the target point cloud, splices the semantic feature of the target point cloud, the target key point feature and the first target point feature to obtain the second target point feature, and finally processes the second target point feature by using the second preset convolution network to obtain the candidate detection box, and determines the target detection box according to the intersection over union between the size of the candidate detection box and the size of the preset box. The point feature processed by convolution has more fine-grained features, the confidence determined according to the point feature has higher accuracy, so that more foreground points can be determined as target key points, and the problem that the generated detection box has low accuracy due to the mixing of too many background points is solved.

[0091] Embodiment three

[0092] Figure 5 is a structural schematic diagram of a detection box determination device provided by the third embodiment of the application. As Figure 5As shown, the device comprises a first feature determination module 301, a confidence determination module 302, a second feature determination module 303, a third feature determination module 304, and a detection frame determination module 305, wherein:

[0093] The first feature determination module is configured to determine a first target point feature and a semantic feature of a target point cloud.

[0094] The confidence determination module is configured to input the first target point feature into a first preset convolutional network, and determine a confidence that each feature in a convolutional point cloud feature belongs to a foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network comprises at least a sparse convolutional layer, a sparse deconvolutional layer, and a popular sparse convolutional layer manifold sparse convolutional layer.

[0095] The second feature determination module is configured to determine an initial key point from the target point cloud, and determine a target key point feature based on the confidence, and distances between the initial key point and each point in the target point cloud except the initial key point.

[0096] The third feature determination module is configured to fuse the semantic feature, the target key point feature, and the first target point feature to obtain a second target point feature.

[0097] The detection frame determination module is configured to process the second target point feature based on a second preset convolutional network to obtain a candidate detection frame, and determine whether to determine the candidate detection frame as a target detection frame according to a difference between a size of the candidate detection frame and a size of a preset frame.

[0098] The determination device of the detection frame provided in the embodiment of the application inputs the first target point feature of the target point cloud into the first preset convolutional network to obtain a confidence that each feature in the convolutional point cloud feature belongs to the foreground point feature, then determines the target key point feature according to the confidence and distances between the initial key point and other points in the target point cloud except the initial key point, then fuses the semantic feature of the target point cloud, the target key point feature, and the first target point feature to obtain the second target point feature, finally processes the second target point feature based on the second preset convolutional network to obtain the candidate detection frame, and determines the target detection frame according to a difference between a size of the candidate detection frame and a size of the preset frame, compared with the traditional detection frame determination method, the three-dimensional space information of the point cloud is retained, the confidence determined by the preset convolutional network can solve the problem that the number of foreground points in the target key point is small, the quality of the foreground point cloud is enhanced, sufficient feature information is provided for subsequent generation of the detection frame, and the accuracy of the detection frame is improved.

[0099] Optionally, the determination manner of the first target point feature comprises: voxelizing an initial point cloud in a preset radar coordinate system to obtain a voxel grid, wherein the voxel grid comprises a plurality of voxel cells; calculating a mean value of coordinates of points in non-empty voxel cells, determining the mean value as a target point coordinate, and generating a plurality of target points according to the target point coordinate, wherein the voxel cells comprise empty voxel cells and the non-empty voxel cells, and each non-empty voxel cell comprises at least one point in the initial point cloud; determining a target point cloud according to the plurality of target points, and determining the target point coordinate as a first target point feature of the target point cloud.

[0100] Optionally, the confidence determination module comprises:

[0101] a first feature determination unit, configured to perform down-sampling processing on the first target point feature by using a sparse convolution layer of a first preset convolution network, to obtain a first feature;

[0102] a second feature determination unit, configured to perform up-sampling processing on the first feature by using a sparse deconvolution layer of the first preset convolution network, to obtain a second feature;

[0103] a confidence determination unit, configured to input the second feature into a popular sparse convolution layer manifold sparse convolution layer of the first preset convolution network, and determine a confidence of each feature in a convolution point cloud feature belonging to a foreground point feature according to an output of a nonlinear classifier in the popular sparse convolution layer manifold sparse convolution layer.

[0104] Optionally, the second feature determination module comprises:

[0105] an initial key point determination unit, configured to determine a first initial key point and a second initial key point from the target point cloud;

[0106] a first distance set determination unit, configured to determine a minimum value of a Euclidean distance between each point in the target point cloud except the initial key points and the initial key points as a first Euclidean distance, to obtain a first Euclidean distance set, wherein the initial key points comprise the first initial key point and the second initial key point;

[0107] a second distance set determination unit, configured to determine a weight coefficient of each first Euclidean distance in the first Euclidean distance set according to a size of the confidence, and calculate a product of the first Euclidean distance and the weight coefficient to obtain a second Euclidean distance set, wherein the weight coefficient is positively correlated with the confidence;

[0108] a key point feature determination unit, configured to, before a quantity of determined target key point features reaches a preset quantity, determine a first target point feature corresponding to a maximum second Euclidean distance in the second Euclidean distance set as a newly added target key point feature.

[0109] The initial value of the number of target key point features is zero, the weight coefficient is adjusted every time a new target key point feature is determined, and the second set of Euclidean distances is updated according to the adjustment of the weight coefficient.

[0110] Optionally, the determining of the first initial key point and the second initial key point from the target point cloud comprises: determining a first initial key point from the target point cloud, and determining a first target point feature corresponding to the first initial key point in the target point cloud as a first initial key point feature; determining a third Euclidean distance between the first initial key point and each point in the target point cloud except the first initial key point according to the first initial key point feature and a first target point feature corresponding to each point in the target point cloud except the first initial key point, to obtain a third set of Euclidean distances; determining a first target point feature corresponding to the maximum third Euclidean distance in the third set of Euclidean distances as a second initial key point feature, and determining a second initial key point according to the second initial key point feature.

[0111] Optionally, the third feature determination module comprises:

[0112] The third feature determination unit is configured to concatenate the target key point feature and the first target point feature to obtain a third feature.

[0113] The target point feature determination unit is configured to down-sample the third feature to a fourth feature matching the dimension of the semantic feature by using a third preset convolutional network, and concatenate the fourth feature and the semantic feature to obtain a second target point feature.

[0114] Optionally, the bounding box determination module comprises:

[0115] The compressed feature determination unit is configured to perform compression processing on the second target point feature to obtain a two-dimensional top view compressed feature.

[0116] The candidate bounding box determination unit is configured to perform clustering processing on the two-dimensional top view compressed feature by using a second preset convolutional network to obtain a candidate bounding box.

[0117] The target bounding box determination unit is configured to determine whether an intersection over union of a size of the candidate bounding box and a size of a preset bounding box is greater than a set threshold value, and if yes, determine the candidate bounding box as a target bounding box.

[0118] The bounding box determination apparatus provided in the embodiments of the present application can perform the bounding box determination method provided in any of the embodiments of the present application, and has the corresponding function modules and beneficial effects of performing the method.

[0119] Embodiment Four

[0120] Figure 6 A structural diagram of an electronic device 40 that can be used to implement embodiments of the present application is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit implementations of the present application described and / or claimed in this document.

[0121] As shown in Figure 6 The electronic device 40 includes at least one processor 41, and a memory, such as a read-only memory (ROM) 42, a random access memory (RAM) 43, etc., connected to the at least one processor 41 in communication, where the memory stores computer programs executable by the at least one processor, and the processor 41 can perform various appropriate actions and processes according to the computer programs stored in the read-only memory (ROM) 42 or loaded into the random access memory (RAM) 43 from the storage unit 48. In the RAM 43, various programs and data required for the operation of the electronic device 40 can also be stored. The processor 41, the ROM 42, and the RAM 43 are connected to each other through a bus 44. An input / output (I / O) interface 45 is also connected to the bus 44.

[0122] Various components in the electronic device 40 are connected to the I / O interface 45, including an input unit 46, such as a keyboard, a mouse, etc., an output unit 47, such as various types of displays, speakers, etc., a storage unit 48, such as a magnetic disk, an optical disk, etc., and a communication unit 49, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 49 allows the electronic device 40 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunications networks.

[0123] The processor 41 can be various general and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 41 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 41 performs various methods and processes described above, such as the determination method of the detection frame.

[0124] In some embodiments, the method of determining a detection box can be implemented as a computer program tangibly embodied in a computer readable storage medium, such as storage unit 48. In some embodiments, part or all of the computer program can be loaded onto and / or installed on electronic device 40 via ROM 42 and / or communication unit 49. One or more steps of the method of determining a detection box described above can be performed when the computer program is loaded into RAM 43 and executed by processor 41. Alternatively, in other embodiments, processor 41 can be configured, by way of firmware or software, to perform the method of determining a detection box.

[0125] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, specially designed application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), a computer hardware, a computer firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0126] Computer programs used to implement the methods of the application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the computer program, when executed by the processor of the machine, implements the functions / acts specified in the flowcharts and / or block diagrams. The computer program can be executed entirely on a machine, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0127] The computer device provided above can be used to execute the method of determining a detection box provided in any of the embodiments above, and has the corresponding functions and advantages.

[0128] Embodiment Five

[0129] In the context of the present application, the computer readable storage medium can be a tangible medium, the computer executable instructions of which, when executed by a computer processor, serve to perform a method of determining a detection box, the method comprising:

[0130] determining a first target point feature and a semantic feature of the target point cloud;

[0131] input the first target point feature into a first preset convolutional network, and determine a confidence degree that each feature in the convolutional point cloud feature belongs to the foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network at least comprises a sparse convolutional layer, a sparse deconvolutional layer, and a popular sparse convolutional layer manifold sparse convolutional layer;

[0132] determine an initial key point from the target point cloud, and determine a target key point feature based on the confidence degree, the initial key point, and a distance between each point in the target point cloud except the initial key point;

[0133] fuse the semantic feature, the target key point feature, and the first target point feature to obtain a second target point feature;

[0134] process the second target point feature based on a second preset convolutional network to obtain a candidate detection frame, and determine whether to determine the candidate detection frame as a target detection frame according to a difference between a size of the candidate detection frame and a size of a preset frame, wherein the second target point feature is a three-dimensional feature.

[0135] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. The computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the above. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of the machine-readable storage medium will include one or more wires, portable computer disks, hard disk drives, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disc read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0136] The computer device provided above can be used to execute the determination method of the detection frame provided in any of the above embodiments, and has the corresponding functions and advantages.

[0137] It is worth noting that in the above embodiment of the determination device of the detection frame, each unit and module is only divided according to the function logic, but is not limited to the above division, as long as the corresponding function can be realized; in addition, the specific name of each functional unit is only for convenient mutual distinction, and does not limit the protection scope of the present application.

[0138] Note that the above merely describes preferred embodiments of the application and the principles of the technology applied. Those skilled in the art will understand that the application is not limited to the specific embodiments described herein, and that various obvious changes, modifications and substitutions can be made to the application without departing from the scope of the application. Therefore, although the application has been described in detail by the above embodiments, the application is not limited to the above embodiments, and can include more other equivalent embodiments without departing from the concept of the application, and the scope of the application is determined by the appended claims.

Claims

1. A method for determining a detection frame, characterized in that, The method comprises the following steps: determining first target point features and semantic features of a target point cloud; inputting the first target point features into a first preset convolutional network, and determining a confidence degree of each feature in a convolutional point cloud feature belonging to a foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network at least comprises a sparse convolutional layer, a sparse deconvolutional layer, and a manifold sparse convolutional layer; determining initial key points from the target point cloud, and determining target key point features based on the confidence degree, distances between the initial key points and each point in the target point cloud except the initial key points; fusing the semantic features, the target key point features, and the first target point features to obtain second target point features; processing the second target point features based on a second preset convolutional network to obtain a candidate detection frame, and determining whether to determine the candidate detection frame as a target detection frame according to a difference between a size of the candidate detection frame and a size of a preset frame, wherein the second target point features are three-dimensional features.

2. The method of claim 1, wherein, The determination manner of the first target point features comprises: voxelizing an initial point cloud in a preset radar coordinate system to obtain a voxel grid, wherein the voxel grid comprises a plurality of voxel cells; calculating a mean value of coordinates of points in non-empty voxel cells, determining the mean value as target point coordinates, and generating a plurality of target points according to the target point coordinates, wherein the voxel cells comprise empty voxel cells and the non-empty voxel cells, and each non-empty voxel cell comprises at least one point in the initial point cloud; determining a target point cloud according to the plurality of target points, and determining the target point coordinates as first target point features of the target point cloud.

3. The method according to claim 1 or 2, characterized in that, The inputting of the first target point features into the first preset convolutional network and the determination of the confidence degree of each feature in the convolutional point cloud feature belonging to the foreground point feature according to the output of the first preset convolutional network comprises: performing down-sampling processing on the first target point features by using a sparse convolutional layer of the first preset convolutional network to obtain first features; performing up-sampling processing on the first features by using a sparse deconvolutional layer of the first preset convolutional network to obtain second features; inputting the second features into a manifold sparse convolutional layer of the first preset convolutional network, and determining the confidence degree of each feature in the convolutional point cloud feature belonging to the foreground point feature according to an output of a non-linear classifier in the manifold sparse convolutional layer.

4. The method of claim 1, wherein, The determination of the target key point features based on the confidence degree, distances between the initial key points and each point in the target point cloud except the initial key points comprises: determining first initial key points and second initial key points from the target point cloud; determining a minimum value of Euclidean distances between each point in the target point cloud except the initial key points and the initial key points as a first Euclidean distance to obtain a first Euclidean distance set, wherein the initial key points comprise the first initial key points and the second initial key points; Determine a weight coefficient of each first Euclidean distance in the first Euclidean distance set according to the size of the confidence, and calculate a product of the first Euclidean distance and the weight coefficient to obtain a second Euclidean distance set, wherein the weight coefficient is positively correlated with the confidence; Before the determined number of target key point features reaches a preset number, determine a first target point feature corresponding to a maximum second Euclidean distance in the second Euclidean distance set as a newly added target key point feature; The initial value of the number of target key point features is zero, and the weight coefficient is adjusted once for each newly added target key point feature, and the second Euclidean distance set is updated according to the adjustment of the weight coefficient.

5. The method of claim 4, wherein, The determining of the first initial key point and the second initial key point from the target point cloud comprises: Determine a first initial key point from the target point cloud, and determine a first target point feature corresponding to the first initial key point in the target point cloud as a first initial key point feature; Determine a third Euclidean distance between the first initial key point and each point in the target point cloud except the first initial key point according to the first initial key point feature and a first target point feature corresponding to each point in the target point cloud except the first initial key point to obtain a third Euclidean distance set; Determine a first target point feature corresponding to a maximum third Euclidean distance in the third Euclidean distance set as a second initial key point feature, and determine a second initial key point according to the second initial key point feature.

6. The method of claim 1, wherein, The fusing of the semantic feature, the target key point feature and the first target point feature to obtain a second target point feature comprises: Splice the target key point feature and the first target point feature to obtain a third feature; Downsample the third feature to a fourth feature matching the dimension of the semantic feature by using a third preset convolutional network, and splice the fourth feature and the semantic feature to obtain a second target point feature.

7. The method of claim 1, wherein, The processing of the second target point feature based on the second preset convolutional network to obtain a candidate bounding box, and determining whether the candidate bounding box is a target bounding box according to a difference between the size of the candidate bounding box and the size of a preset bounding box, comprises: Compress the second target point feature to obtain a two-dimensional top view compressed feature; Cluster the two-dimensional top view compressed feature by using a second preset convolutional network to obtain a candidate bounding box; Determine whether an intersection over union of the size of the candidate bounding box and the size of a preset bounding box is greater than a set threshold, and if so, determine the candidate bounding box as a target bounding box.

8. A detection frame determining apparatus characterized by comprising: Comprise: A first feature determination module configured to determine a first target point feature and a semantic feature of a target point cloud; A confidence determination module configured to input the first target point feature into a first preset convolutional network, and determine a confidence that each feature in a convolutional point cloud feature belongs to a foreground point feature according to an output of the first preset convolutional network, wherein the first preset convolutional network at least comprises a sparse convolutional layer, a sparse deconvolutional layer and a manifold sparse convolutional layer; The second feature determination module is configured to determine an initial key point from the target point cloud, and determine a target key point feature based on the confidence, the initial key point, and a distance between each point in the target point cloud except the initial key point. The third feature determination module is configured to fuse the semantic feature, the target key point feature, and the first target point feature to obtain a second target point feature. The bounding box determination module is configured to process the second target point feature based on a second preset convolutional network to obtain a candidate bounding box, and determine whether to determine the candidate bounding box as a target bounding box according to a difference between a size of the candidate bounding box and a size of a preset box, wherein the second target point feature is a three-dimensional feature.

9. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the method for determining a bounding box according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are used to enable the processor to implement the method for determining a bounding box according to any one of claims 1-7 when executed.

Images