Method and apparatus for training 3D target detection model

Abstract

The application provides a 3D target detection model training method and device, and relates to the technical field of artificial intelligence, wherein the 3D target detection model training method comprises the following steps: obtaining a point cloud data set; inputting the point cloud data set into an initial 3D target detection model to obtain initial pseudo labels corresponding to the point cloud data set; inputting the initial pseudo labels into a local channel for processing to obtain local pseudo labels corresponding to the point cloud data set; inputting the initial pseudo labels into a global channel for processing to obtain global pseudo labels corresponding to the point cloud data set; merging and processing the local pseudo labels and the global pseudo labels to determine target pseudo labels corresponding to the point cloud data set; and training the initial 3D target detection model based on the point cloud data set and the target pseudo labels until a training stop condition is reached to obtain a trained 3D target detection model. Through the above method, the detection accuracy of the 3D target detection model is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a training method and apparatus for a 3D object detection model. Background Technology

[0002] In recent years, with the development of autonomous driving, 3D object detection has received increasing attention from academia and industry. Most 3D-related tasks, such as autonomous driving, utilize LiDAR point cloud data to obtain accurate three-dimensional spatial information of objects.

[0003] In related technologies, 3D object detection models are usually used for 3D object detection. However, since the data used to train 3D object detection models is point cloud data, which is different from previous image data, it has properties such as high dimensionality and disorder, which also causes problems such as difficulty in data annotation, high annotation cost, and large differences in dataset distribution.

[0004] Furthermore, the performance of 3D object detection models degrades significantly when tested on locations, states, or sensor data (target domain) that differ from the training data (source domain). This is due to the differences in the distribution between the source and target domain data. Therefore, improving the detection accuracy of 3D object detection models is a pressing issue that needs to be addressed. Summary of the Invention

[0005] To address the problems existing in the prior art, embodiments of the present invention provide a training method and apparatus for a 3D target detection model.

[0006] This invention provides a training method for a 3D object detection model, comprising:

[0007] Obtain a point cloud dataset for testing an initial 3D object detection model in the target domain; the initial 3D object detection model is trained based on point cloud data samples from the source domain.

[0008] The point cloud dataset is input into the initial 3D object detection model to obtain the initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset;

[0009] The initial pseudo-labels are input into the local pathway for processing to obtain local pseudo-labels corresponding to the point cloud dataset; the local pathway is used to extract local features corresponding to the point cloud dataset.

[0010] The initial pseudo-labels are input into the global path for processing to obtain global pseudo-labels corresponding to the point cloud dataset; the global path is used to extract global features corresponding to the point cloud dataset.

[0011] The local pseudo-labels and the global pseudo-labels are merged to determine the target pseudo-labels corresponding to the point cloud dataset;

[0012] The initial 3D object detection model is trained based on the point cloud dataset and the target pseudo-labels until the training stops, resulting in a trained 3D object detection model.

[0013] Optionally, the step of processing the initial pseudo-labels into local pathways to obtain local pseudo-labels corresponding to the point cloud dataset includes:

[0014] The initial pseudo-label is input into the local memory module in the local path, and the local memory module matches the initial pseudo-label with the first pseudo-label included in the local memory module to obtain the matching result; the local memory module is used to maintain the spatiotemporal consistency in the current coordinate system.

[0015] Based on the matching results, the local pseudo-label is determined.

[0016] Optionally, the step of inputting the initial pseudo-label into the local memory module in the local path, and having the local memory module match the initial pseudo-label with the first pseudo-label included in the local memory module to obtain a matching result, includes:

[0017] The initial pseudo-label is input into the local memory module, which then matches the initial pseudo-label with the first pseudo-label based on the Intersection over Union (IOU) algorithm to obtain a matching result.

[0018] The matching results include a matching subset and a non-matching subset; the matching subset includes an initial pseudo-label and a first pseudo-label whose IOU reaches a first threshold, and the non-matching subset includes an initial pseudo-label and a first pseudo-label whose IOU does not reach the first threshold.

[0019] Optionally, determining the local pseudo-label based on the matching result includes:

[0020] The local memory module is updated based on the matching subset and the non-matching subset;

[0021] The local pseudo-label is determined based on the updated local memory module.

[0022] Optionally, the step of inputting the initial pseudo-labels into the global path for processing to obtain global pseudo-labels corresponding to the point cloud dataset includes:

[0023] The initial pseudo-label is input into the global path to obtain a first initial global pseudo-label; the first initial global pseudo-label is obtained by projecting the initial pseudo-label from the local coordinate system to the global coordinate system in the global path.

[0024] Based on the nonmaximum suppression (NMS) algorithm, a second initial global pseudo-label is obtained from the first initial global pseudo-label; the second initial global pseudo-label is the initial global pseudo-label with the highest confidence among the first initial global pseudo-labels.

[0025] The second initial global pseudo-labels are merged to obtain the merged result;

[0026] Based on the merging results, the global pseudo-label is determined.

[0027] Optionally, the step of merging the second initial global pseudo-labels to obtain the merged result includes:

[0028] For each second initial global pseudo-tag, calculate the IOU between the second initial global pseudo-tag and the other second initial global pseudo-tags; merge the second initial global pseudo-tags whose IOU reaches a second threshold and whose number of tags reaches a third threshold to obtain the merging result;

[0029] or,

[0030] The second initial global pseudo-label is determined as the merging result.

[0031] Optionally, the target pseudo-labels include a first group and a second group; the first group includes target pseudo-labels corresponding to stationary objects, and the second group includes target pseudo-labels corresponding to moving objects;

[0032] The step of merging the local pseudo-labels and the global pseudo-labels to determine the target pseudo-label corresponding to the point cloud dataset includes:

[0033] The local pseudo-labels are projected from the local coordinate system to the global coordinate system, and the IOU between the local pseudo-labels and the global pseudo-labels is calculated. A second pseudo-label and a third pseudo-label are determined from the global pseudo-labels. The second pseudo-label is a global pseudo-label whose IOU with the local pseudo-labels reaches a fourth threshold, and the third pseudo-label is a global pseudo-label whose IOU with the local pseudo-labels does not reach the fourth threshold.

[0034] Project the second pseudo-label and the third pseudo-label from the global coordinate system to the local coordinate system, add the second pseudo-label to the first group, and add the third pseudo-label to the second group.

[0035] Optionally, training the initial 3D object detection model based on the point cloud dataset and the target pseudo-labels until the training stopping condition is met to obtain a trained 3D object detection model includes:

[0036] Based on the point cloud dataset and the target pseudo-labels, the initial 3D target detection model is trained using a target loss function until the target loss function reaches a fifth threshold, and / or the number of training iterations of the initial 3D target detection model reaches a sixth threshold, thus obtaining the trained 3D target detection model; the target loss function includes: a motion-aware loss function and an IOU loss function;

[0037] The motion perception loss function is expressed by formula (1):

[0038]

[0039] Among them, L rpn L represents the motion sensing loss function; reg L represents the regression loss of the candidate boxes; dir Represents the binary cross-entropy angle loss; a s b s r s and a m b m r m This represents the loss weights for the first group and the second group, respectively.

[0040] The present invention also provides a training apparatus for a 3D object detection model, comprising:

[0041] The acquisition module is used to acquire a point cloud dataset for testing an initial 3D object detection model in the target domain; the initial 3D object detection model is trained based on point cloud data samples from the source domain.

[0042] The first input module is used to input the point cloud dataset into the initial 3D object detection model to obtain the initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset;

[0043] The second input module is used to input the initial pseudo-labels into the local pathway for processing to obtain local pseudo-labels corresponding to the point cloud dataset; the local pathway is used to extract local features corresponding to the point cloud dataset.

[0044] The third input module is used to input the initial pseudo-labels into the global path for processing to obtain global pseudo-labels corresponding to the point cloud dataset; the global path is used to extract global features corresponding to the point cloud dataset.

[0045] The merging module is used to merge the local pseudo-labels and the global pseudo-labels to determine the target pseudo-labels corresponding to the point cloud dataset;

[0046] The training module is used to train the initial 3D object detection model based on the point cloud dataset and the target pseudo-labels until the training stops, thus obtaining a trained 3D object detection model.

[0047] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a training method for the 3D object detection model as described above.

[0048] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a training method for a 3D object detection model as described above.

[0049] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a training method for any of the 3D object detection models described above.

[0050] The training method and apparatus for a 3D object detection model provided by this invention inputs a target domain point cloud dataset into an initial 3D object detection model trained based on source domain point cloud data samples, obtaining initial pseudo-labels corresponding to the target domain point cloud dataset. By inputting the initial pseudo-labels into the local and global pathways of a Global-Local DualPathway architecture, higher-precision local and global pseudo-labels are obtained. Then, the local and global pseudo-labels are merged to further improve the accuracy of the pseudo-labels, resulting in target pseudo-labels corresponding to the target domain point cloud dataset. Based on the target pseudo-labels as supervision information, the initial 3D object detection model is iteratively trained, effectively solving the problem of decreased cross-domain inference performance in 3D detection models and improving the cross-domain generalization ability and detection accuracy of the 3D object detection model. Furthermore, training the initial 3D object detection model using target pseudo-labels achieves performance improvement of the trained 3D object detection model without relying on any manually labeled real data, avoiding the need for training data labeling and reducing manual costs. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 This is one of the flowcharts illustrating the training method for the 3D object detection model provided by the present invention;

[0053] Figure 2 This is a schematic diagram illustrating the spatiotemporal differences in prediction results from existing 3D target detection models.

[0054] Figure 3 This is the second flowchart illustrating the training method for the 3D object detection model provided by this invention;

[0055] Figure 4 This is a logical schematic diagram of the training method for the 3D object detection model provided by the present invention;

[0056] Figure 5 This is a schematic diagram of the structure of the training device for the 3D target detection model provided by the present invention;

[0057] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0059] First, it should be noted that the training method and apparatus for the 3D object detection model provided by this invention addresses the difficulties encountered by existing technologies in adaptive 3D object detection in unsupervised domains by constructing a Global-Local Dual Pathway (GLP) architecture, which utilizes the spatiotemporal consistency constraints of objects in three-dimensional space to obtain high-quality target pseudo-labels.

[0060] The local path outputs the local pseudo-labels predicted by the source domain detection model in the current frame coordinate system, while the global path generates accurate global pseudo-labels by utilizing the model prediction results with large self-motion correlation fusion errors of the sensor.

[0061] During the target domain detection model training phase, local pseudo-labels and accurate global pseudo-labels are merged to obtain the final pseudo-label of the current frame for training. During the training process, multiple loss functions are combined to optimize the model, and through multiple rounds of iterative updates, the number of false positive pseudo-labels is reduced, thereby improving the model detection accuracy.

[0062] The following is combined Figures 1 to 4 The training method of the 3D object detection model provided by this invention will be described in detail. Figure 1 This is one of the flowcharts illustrating the training method for the 3D object detection model provided by this invention. See [link / reference]. Figure 1 As shown, the method includes steps 101-106, wherein:

[0063] Step 101: Obtain a point cloud dataset for testing the initial 3D object detection model in the target domain; the initial 3D object detection model is trained based on point cloud data samples from the source domain.

[0064] First, it should be noted that the subject of this invention can be any electronic device capable of training a 3D object detection model, such as a smartphone, smartwatch, desktop computer, laptop, etc.

[0065] Since the performance of a 3D object detection model degrades significantly when tested on locations, states, or sensor data (target domain) that differ from the training data (source domain), this embodiment first requires obtaining a point cloud dataset for testing the initial 3D object detection model in the target domain in order to improve the detection accuracy of the 3D object detection model. The initial 3D object detection model is trained based on point cloud data samples from the source domain.

[0066] It should be noted that the point cloud dataset includes point cloud data from consecutive frames.

[0067] Step 102: Input the point cloud dataset into the initial 3D object detection model to obtain the initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset.

[0068] In this embodiment, in order to make the initial 3D target detection model Φ k To maintain high detection accuracy in the target domain and improve the cross-domain generalization ability of the 3D object detection model, the target domain point cloud dataset needs to be input into the initial 3D object detection model Φ. k This yields the initial pseudo-labels output by the initial 3D object detection model, corresponding to the target domain point cloud dataset.

[0069] Specifically, on the unlabeled target domain point cloud dataset, assume that the current training iteration is the kth round (round 1 refers to training all data once), where k∈,1,N-, and N is the maximum number of iterations.

[0070] For continuous point clouds Where t is the index of the current frame, n is the length of the point cloud sequence, and PC t and T t This represents the point cloud data and pose of the current frame, where the pose is obtained using a point cloud-based Simultaneous Localization and Mapping (SLAM) algorithm.

[0071] *D t +Input initial 3D object detection modelΦ k The output results are processed by non-maximum suppression (NMS) to obtain the initial 3D object detection model Φ. k The output contains the point cloud prediction result for frame t with a larger error (i.e., the initial pseudo-label).

[0072] Where j represents the index of the object detection box in the prediction result; p t It is the number of predicted object detection boxes in the point cloud of frame t; object detection boxes [P j ]By 3D bounding box B j (including orientation angle) and confidence level C j Composition, expressed as [P j ]=,B j C j -

[0073] Step 103: Input the initial pseudo-labels into the local pathway for processing to obtain local pseudo-labels corresponding to the point cloud dataset; the local pathway is used to extract local features corresponding to the point cloud dataset.

[0074] In this embodiment, because the 3D object detection model processes the point cloud sequence frame by frame, the prediction results between adjacent frames often differ significantly. Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the spatiotemporal differences in prediction results of existing 3D target detection models.

[0075] Therefore, in order to improve the 3D object detection model, this invention constructs a global-local dual-path architecture, the purpose of which is to use the spatiotemporal consistency constraints of objects in three-dimensional space to obtain high-quality target pseudo-labels.

[0076] The global-local dual-path architecture includes a local path and a global path. The local path outputs the predicted local pseudo-labels in the current frame coordinate system, while the global path generates accurate global pseudo-labels by using the sensor's self-motion correlation to fuse the model prediction results with large errors.

[0077] In practical applications, the local pathway includes a local memory (LM) module. Initial pseudo-labels are input into the local pathway for processing to update the LM. Based on the updated LM, the local pseudo-label *L corresponding to the target domain point cloud dataset is determined. l +.

[0078] Step 104: Input the initial pseudo-labels into the global path for processing to obtain global pseudo-labels corresponding to the point cloud dataset; the global path is used to extract global features corresponding to the point cloud dataset.

[0079] In this embodiment, the global path includes a global memory module (GM). The initial pseudo-label is input into the global path for processing to update the GM. Then, based on the updated GM, the global pseudo-label *L corresponding to the target domain point cloud dataset is determined. g +. Among them, the global memory module is used to maintain spatiotemporal consistency in the global coordinate system.

[0080] Step 105: Merge the local pseudo-labels and the global pseudo-labels to determine the target pseudo-labels corresponding to the point cloud dataset.

[0081] In this embodiment, local pseudo-labels and global pseudo-labels are merged to further improve the accuracy of pseudo-labels and obtain target pseudo-labels corresponding to the target domain point cloud dataset.

[0082] Step 106: Train the initial 3D object detection model based on the point cloud dataset and the target pseudo-label until the training stops, and obtain the trained 3D object detection model.

[0083] The training method for the 3D object detection model provided by this invention involves inputting a target domain point cloud dataset into an initial 3D object detection model trained based on source domain point cloud data samples, resulting in initial pseudo-labels corresponding to the target domain point cloud dataset. By inputting these initial pseudo-labels into the local and global pathways of a Global-Local Dual Pathway architecture, higher-precision local and global pseudo-labels are obtained. These local and global pseudo-labels are then merged to further improve their accuracy, yielding target pseudo-labels corresponding to the target domain point cloud dataset. Based on these target pseudo-labels as supervisory information, the initial 3D object detection model is iteratively trained, effectively solving the problem of decreased cross-domain inference performance in 3D detection models and improving their cross-domain generalization ability and detection accuracy. Furthermore, training the initial 3D object detection model using target pseudo-labels achieves performance improvements without relying on any manually labeled real data, avoiding the need for training data labeling and reducing manual costs.

[0084] Optionally, the local pathway includes a local memory module LM; the local memory module includes a first pseudo-tag [M]. i ] k-1 ;

[0085] The process of inputting the initial pseudo-labels into the local pathway to determine the local pseudo-labels corresponding to the point cloud dataset can be achieved through the following steps:

[0086] Step 1) Input the initial pseudo-label into the local memory module in the local path, and the local memory module matches the initial pseudo-label with the first pseudo-label included in the local memory module to obtain the matching result; the local memory module is used to maintain the spatiotemporal consistency in the current coordinate system.

[0087] Step 2) Based on the matching results, determine the local pseudo-labels.

[0088] In this embodiment, let LM be... The initial pseudo-tags need to be... With the first pseudo-label [M] i ] k-1 Perform matching to obtain matching results, and then determine the local pseudo-label {L} corresponding to the target domain point cloud dataset based on the matching results. l}

[0089] In the above implementation, by matching the initial pseudo-label with the first pseudo-label through a local path, the accuracy of the initial pseudo-label can be improved to obtain a local pseudo-label, thereby providing a data foundation for generating a more accurate target pseudo-label.

[0090] Optionally, the step of inputting the initial pseudo-label into the local memory module in the local path, and having the local memory module match the initial pseudo-label with the first pseudo-label included in the local memory module to obtain a matching result, can be implemented through the following steps:

[0091] The initial pseudo-label is input into the local memory module, which then matches the initial pseudo-label with the first pseudo-label based on the Intersection over Union (IOU) algorithm to obtain a matching result. The matching result includes a matching subset and a non-matching subset. The matching subset includes the initial pseudo-label and the first pseudo-label whose IOU reaches a first threshold, and the non-matching subset includes the initial pseudo-label and the first pseudo-label whose IOU does not reach the first threshold.

[0092] In this embodiment, it is necessary to set the initial pseudo-label. The first pseudo-label in LM [M i ] k-1 The Intersection over Union (IOU) ratio is calculated for matching to obtain the matching results; the matching results include matching subsets. and non-matching subsets

[0093] It should be noted that, and The IOU reaches the first threshold. and The IOU did not reach the first threshold.

[0094] Optionally, determining the local pseudo-label based on the matching result is specifically achieved through the following steps 1)-2):

[0095] Step 1) Update the local memory module based on the matching subset and the non-matching subset;

[0096] Step 2) Determine the local pseudo-label based on the updated local memory module.

[0097] In this embodiment, after obtaining the matching results, the local memory module needs to be updated based on the matching subset and the non-matching subset.

[0098] Specifically, for the matching subset:

[0099] Retain the detection boxes with higher confidence and delete the detection boxes with lower confidence. Then update the confidence of the detection boxes, specifically expressed by the following formula (2):

[0100]

[0101] in, To match the confidence of the i-th detection box in subset M, γ is a discount factor used to control the confidence of the merged detection boxes, so as to reduce the fluctuation of the model prediction results.

[0102] Matching subsets If a detection box contains an unmatched bounding box, its confidence level is reduced, and the bounding box is removed from the LM when the confidence level falls below a set threshold.

[0103] The reduction of the confidence of the detection box is specifically expressed by the following formula (3):

[0104]

[0105] For non-matching subsets These are typically objects not detected in the first k-1 iterations. To reduce the impact of false positive predictions on the final pseudo-labels, they are directly added to the LM, and their confidence is updated simultaneously. Specifically, this can be expressed by the following formula (4):

[0106]

[0107] Finally, update LM to [M i ] k Based on the updated LM, local pseudo-labels for the k-th round can be generated. Where j represents the index of the object, l t Indicates the quantity of objects.

[0108] In the above implementation, by matching the initial pseudo-label with the first pseudo-label through a local path, the accuracy of the initial pseudo-label can be improved to obtain a local pseudo-label, thereby providing a data foundation for generating a more accurate target pseudo-label.

[0109] Optionally, the step of inputting the initial pseudo-label into the global path for processing to obtain the global pseudo-label corresponding to the point cloud dataset can be achieved through the following steps [1]-[4]:

[0110] Step [1]: Input the initial pseudo-label into the global path to obtain the first initial global pseudo-label; the first initial global pseudo-label is obtained by projecting the initial pseudo-label from the local coordinate system to the global coordinate system in the global path;

[0111] Step [2]: Based on the nonmaximum suppression (NMS) algorithm, obtain the second initial global pseudo-label from the first initial global pseudo-label; the second initial global pseudo-label is the initial global pseudo-label with the highest confidence among the first initial global pseudo-labels;

[0112] Step [3]: Merge the second initial global pseudo-labels to obtain the merged result;

[0113] Step [4]: ​​Based on the merging result, determine the global pseudo-label.

[0114] In this embodiment, the initial pseudo-label is... Input global path, Projecting from the local coordinate system to the global coordinate system yields the first initial global pseudo-label.

[0115] Among them, The projection from the local coordinate system to the global coordinate system can be specifically represented by the following formulas (5)-(6):

[0116]

[0117]

[0118] in, T is the point representation of the j-th target bounding box in the local coordinate system l. t This is the matrix representation of the sensor's pose at time T, and also the transformation matrix from the current local coordinate system to the global coordinate system. The superscript l indicates the local coordinate system, and g indicates the global coordinate system. This represents the rotation matrix of the detection box and the coordinates of its center point. The rotation matrix is ​​derived from the orientation angle of the detection box.

[0119] In After projecting from the local coordinate system to the global coordinate system, it is necessary to use the Non-Maximum Suppression (NMS) algorithm to obtain the second initial global pseudo-label with the highest confidence from the first initial global pseudo-label.

[0120] Since the second initial global pseudo-label is the union of the prediction results of all point cloud sequences, there are many detection boxes with high overlap for stationary objects. Therefore, it is necessary to merge the second initial global pseudo-label to obtain the merged result.

[0121] After obtaining the merge results Afterwards, based on Global pseudo-tags can be identified.

[0122] Specifically, let GM be... use Update GM, and finally generate the global pseudo-tag for round k. Where j is the index of the object, g nThis represents the number of objects in the global coordinate system.

[0123] In the above implementation, by inputting the initial pseudo-label into the global path, the initial pseudo-label is projected from the local coordinate system to the global coordinate system to obtain the first initial global pseudo-label. Then, the second initial global pseudo-label with the highest confidence is obtained from the first initial global pseudo-label. The second initial global pseudo-label is merged, and based on the merging result, a more accurate global pseudo-label is determined, thereby providing a data basis for generating a more accurate target pseudo-label.

[0124] Optionally, the merging process of the second initial global pseudo-labels to obtain the merged result includes at least one of the following methods:

[0125] Method [1]: For each of the second initial global pseudo-labels, calculate the IOU between the second initial global pseudo-label and the other second initial global pseudo-labels; merge the second initial global pseudo-labels whose IOU reaches the second threshold and whose number of labels reaches the third threshold to obtain the merging result.

[0126] In this embodiment, overlapping redundant detection boxes are merged by weighted averaging.

[0127] Specifically, for each second initial global pseudo-label Will Each detection box in the cluster is considered as the cluster center, and then the other detection boxes are calculated. The detection box is added to the cluster when its IOU with the cluster center is greater than the second threshold (e.g., 0.5). Finally, clusters with a number of detection boxes reaching the third threshold (e.g., 5) are merged, specifically represented by the following formulas (7) and (8):

[0128]

[0129]

[0130] in, It is a detection box The confidence score of the merged bounding box among all adjacent bounding boxes in the cluster. Set to highest confidence level The final output is the accurate global detection box (i.e., the merged result).

[0131] Method [2]: The second initial global pseudo-label is determined as the merging result. That is...

[0132] In the above implementation, by merging the second initial global pseudo-labels, the global temporal consistency of static targets is ensured, thereby generating global pseudo-labels with spatiotemporal consistency, which in turn provides a data foundation for generating more accurate target pseudo-labels.

[0133] Optionally, after determining the local pseudo-labels and global pseudo-labels corresponding to the target domain point cloud dataset, the local pseudo-labels and global pseudo-labels need to be merged to determine the target pseudo-labels corresponding to the target domain point cloud dataset.

[0134] Among them, the target pseudo-labels include the first group O s Second group O m The first group includes target pseudo-labels corresponding to stationary objects, and the second group includes target pseudo-labels corresponding to moving objects.

[0135] It should be noted that in the process of merging local pseudo-labels and global pseudo-labels, the global pseudo-labels in the global coordinate system need to be back-projected back to the local coordinate system first. Then, the pseudo-labels visible in the current local coordinate system are selected and merged with the local pseudo-labels to generate the final pseudo-labels used for training.

[0136] The process of merging the local pseudo-labels and the global pseudo-labels to determine the target pseudo-labels corresponding to the target domain point cloud dataset can be achieved through the following steps (1)-(2):

[0137] Step (1): Project the local pseudo-label from the local coordinate system to the global coordinate system, calculate the IOU between the local pseudo-label and the global pseudo-label, and determine the second pseudo-label and the third pseudo-label from the global pseudo-label; the second pseudo-label is a global pseudo-label whose IOU with the local pseudo-label reaches the fourth threshold, and the third pseudo-label is a global pseudo-label whose IOU with the local pseudo-label does not reach the fourth threshold.

[0138] Step (2): Project the second pseudo-label and the third pseudo-label from the global coordinate system to the local coordinate system, add the second pseudo-label to the first group, and add the third pseudo-label to the second group.

[0139] In this embodiment, the problem of selecting visible pseudo-labels in the current local coordinate system from global pseudo-labels can be transformed into a retrieval problem, i.e., using the local pseudo-labels output by LM. As a query set, retrieve the corresponding pseudo-tag from the global pseudo-tags.

[0140] Specifically, local pseudo-tags Projecting from the local coordinate system to the global coordinate system, calculate and global pseudo tags If a detection box with an IOU exceeding the fourth threshold (e.g., 0.5) is matched, the matched detection box is back-projected back to the local coordinate system and added to the target pseudo-label set.

[0141] It should be noted that for the detection bounding boxes of stationary objects, the corresponding global pseudo-labels can usually be matched. Therefore, the back-projected detection bounding boxes are also added to the stationary object group (i.e., the first group). s middle.

[0142] If no detection box is found, the detection box is determined to be a moving object, and it is directly added to the target pseudo-label set, and simultaneously added to the moving object group (i.e., the second group). m middle.

[0143] Among them, local pseudo tags The projection from the local coordinate system to the global coordinate system can be achieved using the following formula (9):

[0144]

[0145] Among them, T t Let be the transformation matrix that projects the point cloud of frame t onto the global coordinate system; These are local pseudo-tags used as part of the query set. The retrieval process returns a global set of pseudo-tags. The detection box that is closest to the query pseudo-label.

[0146] Furthermore, in order to reduce the number of pseudo-labels that are semantically ambiguous, the triplet box partition method in ST3D can be used to further filter the target pseudo-labels, specifically expressed by the following formula (10):

[0147]

[0148] Where, σ ignore and σ positive It is the threshold for determining positive and ignored tags; S i This indicates the supervision status of the pseudo-label.

[0149] For any input sample and its pseudo-label L j If S i If S = 0, then the sample is a negative sample; if S i =-1, then ignore this sample; if S i If the value is 1, then it is a positive sample.

[0150] Optionally, in one possible implementation of this invention, after merging the local and global pseudo-labels to determine the target pseudo-label corresponding to the target domain point cloud dataset, the initial 3D object detection model can be trained based on the target domain point cloud dataset and the target pseudo-label until the training stopping condition is met, resulting in a trained 3D object detection model. Specifically, this can be achieved in the following ways:

[0151] Based on the point cloud dataset and the target pseudo-label, the initial 3D target detection model is trained using the target loss function until the target loss function reaches the fifth threshold and / or the number of training iterations of the initial 3D target detection model reaches the sixth threshold, thus obtaining the trained 3D target detection model.

[0152] In this embodiment, the target loss function includes: a motion-aware loss function and an IOU loss function, i.e., L det =L rpn +L iou Among them, L det Let L represent the target loss function. rpn L represents the motion-sensing loss function. iou This represents the IOU loss function.

[0153] It should be noted that the motion perception loss function is expressed by formula (1):

[0154]

[0155] Among them, L rpn L represents the motion sensing loss function; reg L represents the regression loss of the candidate boxes; dir This represents the binary cross-entropy angle loss, used to estimate the orientation angle of candidate boxes; a s b s r s and a m b m r m This represents the respective loss weights for the first and second groups. For the moving object group (i.e., the second group), the weight of the classification loss is increased to improve classification accuracy; for the stationary object group (i.e., the first group), the weights of the regression loss and angle loss are increased to improve positioning accuracy.

[0156] By using the above method, the initial 3D object detection model is iteratively trained based on the target pseudo-label as supervision information. This effectively solves the problem of declining cross-domain inference performance of the 3D detection model and improves the cross-domain generalization ability and detection accuracy of the 3D object detection model.

[0157] Figure 3This is the second flowchart illustrating the training method for the 3D object detection model provided by this invention. See also... Figure 3 As shown, the method includes steps 301-312, wherein:

[0158] Step 301: Obtain the point cloud dataset for testing the initial 3D object detection model in the target domain; wherein, the initial 3D object detection model is trained based on point cloud data samples from the source domain.

[0159] Step 302: Input the point cloud dataset into the initial 3D object detection model to obtain the initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset.

[0160] Step 303: Input the initial pseudo-label into the local memory module in the local pathway. The local memory module matches the initial pseudo-label with the first pseudo-label based on the IOU algorithm to obtain the matching result. The matching result includes a matching subset and a non-matching subset. The matching subset includes the initial pseudo-label and the first pseudo-label with IOU reaching the first threshold, and the non-matching subset includes the initial pseudo-label and the first pseudo-label with IOU not reaching the first threshold.

[0161] Step 304: Update the local memory module based on the matching subset and the non-matching subset; determine the local pseudo-label based on the updated local memory module.

[0162] Step 305: Input the initial pseudo-label into the global path to obtain the first initial global pseudo-label; wherein, the first initial global pseudo-label is obtained by projecting the initial pseudo-label from the local coordinate system to the global coordinate system in the global path.

[0163] Step 306: Based on the NMS algorithm, obtain the second initial global pseudo-label from the first initial global pseudo-label; wherein the second initial global pseudo-label is the initial global pseudo-label with the highest confidence among the first initial global pseudo-labels.

[0164] Step 307: For each second initial global pseudo-label, calculate the IOU between the second initial global pseudo-label and other second initial global pseudo-labels; merge the second initial global pseudo-labels whose IOU reaches the second threshold and whose number of labels reaches the third threshold to obtain the merged result; or, determine the second initial global pseudo-label as the merged result.

[0165] Step 308: Based on the merging results, determine the global pseudo-label.

[0166] Step 309: Merge the local pseudo-labels and global pseudo-labels to determine the target pseudo-labels corresponding to the point cloud dataset.

[0167] Step 310: Based on the target domain point cloud dataset and target pseudo-labels, train the initial 3D target detection model using the target loss function.

[0168] Step 311: Determine whether the training stop condition has been met; if yes, proceed to step 312; if no, return to step 302.

[0169] Specifically, the training stopping conditions are: the target loss function reaches the fifth threshold, and / or the initial 3D object detection model has been trained for a sixth time. The target loss function includes the motion-aware loss function and the IOU loss function.

[0170] Step 312: Obtain the trained 3D object detection model.

[0171] Figure 4 This is a logical schematic diagram of the training method for the 3D object detection model provided by the present invention; as shown below. Figure 4 As shown,

[0172] Step 1: Convert the point cloud sequence Input to the initial 3D object detection model Φ k The prediction results are obtained (i.e., the initial pseudo-labels mentioned above). ).

[0173] Step 2, By using local paths (i.e., the local pathways mentioned above), a local coordinate system is obtained (i.e., ), and update the local memory module to obtain the local pseudo-label.

[0174] Step 3, The global coordinate system (i.e., the global path mentioned above) is obtained through the global path. ), and update the global memory module to obtain the global pseudo-tag.

[0175] Step 4: Remove local pseudo-tags and global pseudo tags The fusion process is performed to obtain target pseudo-labels, which include groups of moving objects (i.e., the second group mentioned above) and groups of stationary objects (i.e., the first group mentioned above).

[0176] Step 5: Apply the motion-aware loss function to the initial 3D object detection model Φ k Training is performed to obtain the 3D object detection model Φ in the (k+1)th round. k+1 .

[0177] Step 6: Let k = k + 1, and repeat steps 1 through 5 until the 3D object detection model converges or the number of iterations k = N + 1. Output the trained, high-performance 3D object detection model Φ. t (i.e., a trained 3D object detection model).

[0178] The training apparatus for the 3D object detection model provided by the present invention will be described below. The training apparatus for the 3D object detection model described below can be referred to in correspondence with the training method for the 3D object detection model described above. Figure 5 This is a schematic diagram of the structure of the training device for the 3D object detection model provided by the present invention, as shown below. Figure 5 As shown, the training device 500 for the 3D object detection model includes: an acquisition module 501, a first input module 502, a second input module 503, a third input module 504, a merging module 505, and a training module 506, wherein:

[0179] The acquisition module 501 is used to acquire a point cloud dataset for testing the initial 3D object detection model in the target domain; the initial 3D object detection model is trained based on point cloud data samples from the source domain.

[0180] The first input module 502 is used to input the point cloud dataset into the initial 3D object detection model to obtain the initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset;

[0181] The second input module 503 is used to input the initial pseudo-label into the local path for processing to obtain a local pseudo-label corresponding to the point cloud dataset; the local path is used to extract local features corresponding to the point cloud dataset.

[0182] The third input module 504 is used to input the initial pseudo-label into the global path for processing to obtain a global pseudo-label corresponding to the point cloud dataset; the global path is used to extract global features corresponding to the point cloud dataset.

[0183] The merging module 505 is used to merge the local pseudo-labels and the global pseudo-labels to determine the target pseudo-labels corresponding to the point cloud dataset.

[0184] The training module 506 is used to train the initial 3D object detection model based on the point cloud dataset and the target pseudo-labels until the training stopping condition is met, so as to obtain the trained 3D object detection model.

[0185] The training device for the 3D object detection model provided by this invention inputs the target domain point cloud dataset into an initial 3D object detection model trained based on source domain point cloud data samples, obtaining initial pseudo-labels corresponding to the target domain point cloud dataset. By inputting the initial pseudo-labels into the local and global pathways of a global-local dual-pathway architecture, higher-precision local and global pseudo-labels are obtained. Then, the local and global pseudo-labels are merged to further improve the accuracy of the pseudo-labels, resulting in target pseudo-labels corresponding to the target domain point cloud dataset. Based on the target pseudo-labels as supervision information, the initial 3D object detection model is iteratively trained, effectively solving the problem of decreased cross-domain inference performance in 3D detection models and improving the cross-domain generalization ability and detection accuracy of the 3D object detection model. Furthermore, training the initial 3D object detection model using target pseudo-labels achieves performance improvement of the trained 3D object detection model without relying on any manually labeled real data, avoiding the need for training data labeling and reducing manual costs.

[0186] Optionally, the first input module 502 is further configured to:

[0187] The initial pseudo-label is input into the local memory module in the local path, and the local memory module matches the initial pseudo-label with the first pseudo-label included in the local memory module to obtain the matching result; the local memory module is used to maintain the spatiotemporal consistency in the current coordinate system.

[0188] Based on the matching results, the local pseudo-label is determined.

[0189] Optionally, the first input module 503 is further configured to:

[0190] The initial pseudo-label is input into the local memory module, which then matches the initial pseudo-label with the first pseudo-label based on the Intersection over Union (IOU) algorithm to obtain a matching result.

[0191] The matching results include a matching subset and a non-matching subset; the matching subset includes an initial pseudo-label and a first pseudo-label whose IOU reaches a first threshold, and the non-matching subset includes an initial pseudo-label and a first pseudo-label whose IOU does not reach the first threshold.

[0192] Optionally, the first input module 503 is further configured to:

[0193] The local memory module is updated based on the matching subset and the non-matching subset;

[0194] The local pseudo-label is determined based on the updated local memory module.

[0195] Optionally, the second input module 504 is further configured to:

[0196] The initial pseudo-label is input into the global path to obtain a first initial global pseudo-label; the first initial global pseudo-label is obtained by projecting the initial pseudo-label from the local coordinate system to the global coordinate system in the global path.

[0197] Based on the nonmaximum suppression (NMS) algorithm, a second initial global pseudo-label is obtained from the first initial global pseudo-label; the second initial global pseudo-label is the initial global pseudo-label with the highest confidence among the first initial global pseudo-labels.

[0198] The second initial global pseudo-labels are merged to obtain the merged result;

[0199] Based on the merging results, the global pseudo-label is determined.

[0200] Optionally, the second input module 504 is further configured to:

[0201] For each second initial global pseudo-tag, calculate the IOU between the second initial global pseudo-tag and the other second initial global pseudo-tags; merge the second initial global pseudo-tags whose IOU reaches a second threshold and whose number of tags reaches a third threshold to obtain the merging result;

[0202] or,

[0203] The second initial global pseudo-label is determined as the merging result.

[0204] Optionally, the target pseudo-labels include a first group and a second group; the first group includes target pseudo-labels corresponding to stationary objects, and the second group includes target pseudo-labels corresponding to moving objects;

[0205] The merging module 505 is further configured to:

[0206] The local pseudo-labels are projected from the local coordinate system to the global coordinate system, and the IOU between the local pseudo-labels and the global pseudo-labels is calculated. A second pseudo-label and a third pseudo-label are determined from the global pseudo-labels. The second pseudo-label is a global pseudo-label whose IOU with the local pseudo-labels reaches a fourth threshold, and the third pseudo-label is a global pseudo-label whose IOU with the local pseudo-labels does not reach the fourth threshold.

[0207] Project the second pseudo-label and the third pseudo-label from the global coordinate system to the local coordinate system, add the second pseudo-label to the first group, and add the third pseudo-label to the second group.

[0208] Optionally, the training module 506 is further configured to:

[0209] Based on the point cloud dataset and the target pseudo-labels, the initial 3D target detection model is trained using a target loss function until the target loss function reaches a fifth threshold, and / or the number of training iterations of the initial 3D target detection model reaches a sixth threshold, thus obtaining the trained 3D target detection model; the target loss function includes: a motion-aware loss function and an IOU loss function;

[0210] The motion perception loss function is expressed by formula (1):

[0211]

[0212] Among them, L rpn L represents the motion sensing loss function; reg L represents the regression loss of the candidate boxes; dir Represents the binary cross-entropy angle loss; a s b s r s and a m b m r m This represents the loss weights for the first group and the second group, respectively.

[0213] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a training method for a 3D object detection model. This method includes: acquiring a point cloud dataset for testing an initial 3D object detection model in a target domain; the initial 3D object detection model is trained based on point cloud data samples from the source domain; inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; inputting the initial pseudo-labels into a local pathway for processing to obtain local pseudo-labels corresponding to the point cloud dataset; the local pathway is used to extract local features corresponding to the point cloud dataset; inputting the initial pseudo-labels into a global pathway for processing to obtain global pseudo-labels corresponding to the point cloud dataset; the global pathway is used to extract global features corresponding to the point cloud dataset; merging the local pseudo-labels and the global pseudo-labels to determine target pseudo-labels corresponding to the point cloud dataset; and training the initial 3D object detection model based on the point cloud dataset and the target pseudo-labels until a training stopping condition is met to obtain a trained 3D object detection model.

[0214] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0215] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute the training method of the 3D object detection model provided by the above methods, the method comprising: acquiring a point cloud dataset for testing an initial 3D object detection model in a target domain; the initial 3D object detection model being trained based on point cloud data samples from a source domain; inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; and inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; and inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; and inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; and inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; and inputting the initial pseudo-labels ... the initial pseudo-labels output by the initial pseudo-labels output by the initial pseudo-labels output by the initial pseudo-labels output by the initial pseudo-labels output by the initial pseudo-label Initial pseudo-labels are input into a local pathway for processing to obtain local pseudo-labels corresponding to the point cloud dataset; the local pathway is used to extract local features corresponding to the point cloud dataset; the initial pseudo-labels are input into a global pathway for processing to obtain global pseudo-labels corresponding to the point cloud dataset; the global pathway is used to extract global features corresponding to the point cloud dataset; the local pseudo-labels and the global pseudo-labels are merged to determine the target pseudo-label corresponding to the point cloud dataset; the initial 3D object detection model is trained based on the point cloud dataset and the target pseudo-label until the training stopping condition is met to obtain a trained 3D object detection model.

[0216] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for training a 3D object detection model provided by the methods described above. This method includes: acquiring a point cloud dataset for testing an initial 3D object detection model in a target domain; the initial 3D object detection model being trained based on point cloud data samples from a source domain; inputting the point cloud dataset into the initial 3D object detection model to obtain initial pseudo-labels output by the initial 3D object detection model corresponding to the point cloud dataset; and inputting the initial pseudo-labels into a local access path for processing. The initial pseudo-labels are processed by a global path to obtain local pseudo-labels corresponding to the point cloud dataset. The local path is used to extract local features corresponding to the point cloud dataset. The initial pseudo-labels are then input into a global path for processing to obtain global pseudo-labels corresponding to the point cloud dataset. The global path is used to extract global features corresponding to the point cloud dataset. The local and global pseudo-labels are merged to determine the target pseudo-label corresponding to the point cloud dataset. The initial 3D object detection model is trained based on the point cloud dataset and the target pseudo-label until the training stops, resulting in a trained 3D object detection model.

[0217] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0218] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0219] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a 3D object detection model, characterized in that, The method comprises: acquiring a point cloud data set for testing an initial 3D target detection model in a target domain; the initial 3D target detection model is obtained by training based on a point cloud data sample of a source domain; inputting the point cloud data set into the initial 3D target detection model to obtain initial pseudo labels corresponding to the point cloud data set output by the initial 3D target detection model; inputting the initial pseudo labels into a local channel for processing to obtain local pseudo labels corresponding to the point cloud data set; the local channel is used for extracting local features corresponding to the point cloud data set; inputting the initial pseudo labels into a global channel for processing to obtain global pseudo labels corresponding to the point cloud data set; the global channel is used for extracting global features corresponding to the point cloud data set; merging and processing the local pseudo labels and the global pseudo labels to determine target pseudo labels corresponding to the point cloud data set; training the initial 3D target detection model based on the point cloud data set and the target pseudo labels until a training stop condition is reached to obtain a trained 3D target detection model. 2.The method of claim 1, wherein, The method comprises: inputting the initial pseudo labels into a local memory module in the local channel, matching the initial pseudo labels with first pseudo labels included in the local memory module by the local memory module to obtain a matching result; the local memory module is used for maintaining spatiotemporal consistency in a current coordinate system; determining the local pseudo labels based on the matching result.

3. The method of claim 2, wherein the 3D object detection model is trained by: The method comprises: inputting the initial pseudo labels into the local memory module, matching the initial pseudo labels with the first pseudo labels by the local memory module based on a parallel ratio IOU algorithm to obtain a matching result; wherein the matching result comprises a matching subset and a non-matching subset; the matching subset comprises initial pseudo labels and first pseudo labels with an IOU reaching a first threshold, and the non-matching subset comprises initial pseudo labels and first pseudo labels with an IOU not reaching the first threshold.

4. The method of claim 3, wherein the 3D object detection model is trained by: The method comprises: updating the local memory module based on the matching subset and the non-matching subset; determining the local pseudo labels based on the updated local memory module. 5.The method of claim 1, wherein, The method comprises: inputting the initial pseudo labels into the global channel to obtain first initial global pseudo labels; the first initial global pseudo labels are obtained by projecting the initial pseudo labels from a local coordinate system to a global coordinate system in the global channel; obtain a second initial global pseudo label from the first initial global pseudo label based on a non-maximum suppression (NMS) algorithm; the second initial global pseudo label is an initial global pseudo label with the highest confidence in the first initial global pseudo label; perform merging processing on the second initial global pseudo label to obtain a merging result; determine the global pseudo label based on the merging result. 6.The method of claim 5, wherein, The merging processing on the second initial global pseudo label to obtain a merging result comprises: For each of the second initial global pseudo label, calculate the IOU of the second initial global pseudo label and other second initial global pseudo labels; perform merging processing on the second initial global pseudo label whose IOU reaches a second threshold and whose label number reaches a third threshold to obtain the merging result; or The second initial global pseudo label is determined as the merging result. The target pseudo label comprises a first group and a second group; the first group comprises a target pseudo label corresponding to a stationary object, and the second group comprises a target pseudo label corresponding to a moving object; 7.The method of claim 1, wherein, The merging processing on the local pseudo label and the global pseudo label to determine a target pseudo label corresponding to the point cloud data set comprises: project the local pseudo label from a local coordinate system to a global coordinate system, calculate the IOU of the local pseudo label and the global pseudo label, and determine a second pseudo label and a third pseudo label from the global pseudo label; the second pseudo label is a global pseudo label whose IOU with the local pseudo label reaches a fourth threshold, and the third pseudo label is a global pseudo label whose IOU with the local pseudo label does not reach the fourth threshold; project the second pseudo label and the third pseudo label from the global coordinate system to the local coordinate system, add the second pseudo label to the first group, and add the third pseudo label to the second group. The training of the initial 3D target detection model based on the point cloud data set and the target pseudo label until a training stop condition is reached to obtain a trained 3D target detection model comprises: 8.The method of claim 1, wherein, train the initial 3D target detection model based on the point cloud data set and the target pseudo label using a target loss function until the target loss function reaches a fifth threshold and / or the number of training of the initial 3D target detection model reaches a sixth threshold to obtain the trained 3D target detection model; the target loss function comprises a motion perception loss function and an IOU loss function; The motion perception loss function is represented by formula (1): comprises: wherein, L rpn represents the motion-aware loss function; L reg represents the regression loss of the candidate box; L dir represents the binary cross-entropy angular loss; a s , b s , r s and a m , b m , r m represent the loss weights of the first group and the second group respectively; O s represents the first group, O m represents the second group. 9.A device for training a 3D object detection model, characterized in that, an acquisition module configured to acquire a point cloud data set used for testing an initial 3D target detection model in a target domain; The initial 3D target detection model is obtained by training based on a point cloud data sample in a source domain; a first input module configured to input the point cloud data set into the initial 3D target detection model to obtain an initial pseudo label corresponding to the point cloud data set output by the initial 3D target detection model; ​ The second input module is configured to input the initial pseudo label into a local path for processing to obtain a local pseudo label corresponding to the point cloud dataset; the local path is configured to extract local features corresponding to the point cloud dataset; The third input module is configured to input the initial pseudo label into a global path for processing to obtain a global pseudo label corresponding to the point cloud dataset; The global path is configured to extract global features corresponding to the point cloud dataset; The merging module is configured to merge and process the local pseudo label and the global pseudo label to determine a target pseudo label corresponding to the point cloud dataset; The training module is configured to train the initial 3D target detection model based on the point cloud dataset and the target pseudo label until a training stop condition is reached to obtain a trained 3D target detection model.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the training method of the 3D target detection model according to any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the training method of the 3D target detection model according to any one of claims 1 to 8.

12. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the training method of the 3D target detection model according to any one of claims 1 to 8.

Images