A three-dimensional target detection method with positioning accuracy prediction

Abstract

The application discloses a three-dimensional target detection method with positioning accuracy prediction, first, the point cloud is input into a point feature extraction backbone network, and local features are gradually aggregated through three set abstraction layers. Then, candidate center points and their features are obtained through a candidate generation layer. Subsequently, the feature vectors are input into a detection head, and classification scores, bounding boxes and positioning accuracy of the bounding boxes are predicted for each candidate center point. Finally, the candidate boxes are post-processed to obtain the final prediction result. The difficulty adaptive positioning accuracy prediction method disclosed by the application aims to predict appropriate positioning accuracy for the target box to solve the problem of more false detections of the single-stage target detector. In three-dimensional target detection, compared with the classic detection head, the multi-task detection head with the positioning accuracy prediction branch is added in the application, and by combining the positioning accuracy prediction value and the classification score prediction value as a new NMS sorting index, the positioning accuracy can be effectively improved and the false detection can be reduced.

Description

Technical Field

[0001] This invention relates to a 3D target detection technology with positioning accuracy prediction, belonging to the field of computer vision technology. Background Technology

[0002] 3D object detection is an upstream processing step in some 3D understanding tasks, and its performance directly affects the results of 3D object segmentation, recognition, and tracking. Given sensor input, the task of 3D object detection is to classify and locate objects in 3D space, with each object represented by a class label with a probability score and a 3D bounding box. In intelligent driving systems, mainstream sensor solutions include LiDAR, cameras, and millimeter-wave radar (RADAR). Depending on the sensor used, 3D object detection methods can be divided into monocular-based, binocular-based, point cloud-based, and multimodal fusion-based methods. 3DSSD proposes an efficient single-stage 3D object detector. To achieve higher detection speed, 3DSSD removes the feature propagation layer and thinning module, utilizes a fusion sampling strategy to retain more foreground points, and adds a candidate generation layer to make target localization more accurate. DA-3DSSD improves the feature extraction backbone network based on 3DSSD, proposing a density-sensitive farthest point sampling strategy and a center density attention module. Analysis of the detection results of 3DSSD and DA-3DSSD algorithms revealed that DA-3DSSD increased the number of positive detections at various difficulty levels, which means it improved the recall rate. However, the number of false positives still accounted for 25% to 45% of the total number of detection results, which resulted in low accuracy.

[0003] 3D object detection methods can generally be roughly divided into two parts: the backbone network and the detection head. The detection head is the collective term for the different branches that branch out from the input point cloud after it has passed through a shared feature extraction network, depending on the target task. The most direct goals of detection are classification and localization, so common detection heads only contain classification and localization branches. However, by adding appropriate task branches, the feature content learned by the backbone network can be adjusted during backpropagation, and the network's prediction results can be further processed using the newly added information during forward propagation. By adding a localization accuracy prediction branch, a new difficulty-adaptive localization accuracy metric, HIoU, is proposed. This not only supervises the network to learn to pay more attention to localization information, but also effectively improves localization accuracy and reduces false detections by combining the localization accuracy prediction value with the classification score prediction value as a new NMS ranking metric. This is the main content of this patent. Summary of the Invention

[0004] Technical issues:

[0005] The present invention aims to provide a three-dimensional target detection method with positioning accuracy prediction, thereby solving the problems of mismatch between confidence level and positioning accuracy and a large number of false detections.

[0006] Technical solution:

[0007] To solve the above problems, the present invention adopts the following technical solution:

[0008] A three-dimensional target detection method with positioning accuracy prediction, the method includes the following steps:

[0009] S1. Preprocess the input raw radar point cloud;

[0010] S2. Input the preprocessed point cloud into the constructed point cloud feature extraction backbone network to obtain the predicted candidate target center points and corresponding feature vectors.

[0011] S3. Construct a multi-task branch detection head, and predict the classification score, bounding box and localization accuracy for each candidate target center point through classification, localization and localization accuracy prediction branches;

[0012] S4. The candidate bounding boxes are filtered through a post-processing algorithm to obtain the final prediction result;

[0013] S5. Construct a loss function to optimize network parameters during training;

[0014] The three-dimensional target detection method with positioning accuracy prediction described above, the specific method for preprocessing the input original radar point cloud in step S1 is as follows: the point cloud is cropped according to the range of the two-dimensional image, retaining the label of the current category and ignoring the labels of other categories; the target point cloud of the current category in each point cloud is saved separately for subsequent data augmentation; simple (0) or complex (1) labels are assigned to the target according to the target's occlusion, truncation, height, etc., for training of the prediction branch of the positioning accuracy label HIoU with difficulty adaptation.

[0015] The three-dimensional target detection method with positioning accuracy prediction, in step S2, involves the main steps of the feature extraction backbone network for the preprocessed point cloud, including the SA layer to gather local features and the CG layer to obtain the candidate center point positions and features.

[0016] The specific method for aggregating local features using the SA layer is as follows: the point cloud is input into the point feature extraction backbone network, and local features are aggregated step by step through three SA layers. Each SA layer first samples the points, then groups the sampled points and performs feature aggregation to extract the features around the sampled points. The sampling method combines density-aware point sampling (DPS) and feature-farthest point sampling (F-FPS) to retain more internal points for sparse targets.

[0017] The specific method for obtaining candidate center point locations and features in the CG layer is as follows: Features extracted after three SA layers are used in the subsequent CG layer to predict the offset between the sampled points obtained from the F-FPS and the instance center points. Then, offsets are added to the sampled points to obtain candidate points. These candidate points are considered center points, and features are aggregated from the surrounding points of the last SA layer to calculate the center point density. Simultaneously, weights learned from the center point density using a multilayer perceptron (MLP) are used to reweight the features, resulting in the final candidate center points and features.

[0018] The three-dimensional target detection method with positioning accuracy prediction is described above. In step S3, the multi-task detection head consists of a classification branch, a positioning branch, and a positioning accuracy prediction branch. The classification branch is responsible for predicting the target classification score, the positioning branch is responsible for predicting the target bounding box, and the positioning accuracy prediction branch is responsible for predicting the positioning accuracy of the target bounding box predicted by the positioning branch.

[0019] The specific structure of the classification branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes, and then through another MLP with an output node of 1, representing the predicted score of the target belonging to the given category (here referring to Car). Since the probability is a value between 0 and 1, the predicted value is then mapped to the range (0,1) by the Sigmoid function, thus obtaining the final classification confidence.

[0020] The specific structure of the localization branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes, then through an MLP with 30 output nodes, where 6 nodes output offset and size, 12 nodes output angle category, and 12 nodes output angle offset. Because the positional differences between targets are too large, the labels used in regression are encoded from the true labels to ensure network convergence. The predicted values ​​of the regression branch also need to be decoded to obtain the true positional information. The predicted values ​​include offset and related information for length, width, and height. The decoding formula is:

[0021]

[0022]

[0023] Among them, (x base ,y base ,z base The coordinates of the candidate center point are shown in the diagram. Orientation estimation uses a bin-based angle regression algorithm. The angle is divided into n intervals (n = 12 in this network), and the probability of the target angle belonging to a certain interval and the offset based on that interval are predicted. The network's prediction result (θ) cls ,θ res The actual target angle is obtained by decoding using the following formula:

[0024]

[0025] The specific structure of the localization branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes in the localization accuracy prediction branch, and then through an MLP with an output node of 1, used to predict the localization accuracy of the target box predicted by the localization branch. To achieve prediction, the location information and labels obtained from the localization branch are first converted into bounding boxes represented by 8 vertex coordinates. Then, by comparing the IoU, GIoU, and the proposed HIoU—three localization accuracy metrics—between the predicted box and the ground truth box, HIoU is selected as the target ground truth for localization accuracy. Finally, under the supervision of the HIoU ground truth, difficulty-adaptive localization accuracy prediction is achieved. Because the prediction results of the regression branch are involved in the calculation of the ground truth, it is important to prevent gradient backpropagation to the regression branch during training. The specific formula for calculating HIoU is:

[0026]

[0027]

[0028]

[0029] Here, "label" refers to the difficulty level label assigned during the preprocessing stage.

[0030] The 3D target detection method with positioning accuracy prediction, specifically the post-processing algorithm for filtering candidate boxes in step S4, involves using the candidate boxes and the predicted HIoU (Highest Hierarchical Value) score multiplied by their classification scores as input. The non-maximum suppression algorithm sorts the input boxes by their scores, continuously selecting the highest-scoring boxes and suppressing those with excessive overlap, ultimately obtaining a series of non-overlapping boxes in 3D. Finally, candidate boxes with scores below a set threshold are removed from the detection result set.

[0031] The specific method for constructing a loss function to optimize and train network parameters in step S5 of the aforementioned 3D target detection method with positioning accuracy prediction is as follows:

[0032] The classification branch is learned under the supervision of continuous center-ness labels. The model learns network parameters by minimizing the cross-entropy loss between the predicted confidence score and the center-ness label, as shown in the formula:

[0033] L1(s,l ctrness )=-(l ctrness log s+(1-l ctrness log(1-s))

[0034]

[0035] Where s represents the predicted score of the candidate centroid. The second formula is used to calculate the centroid label of the candidate centroid, where (f,b,l,r,t,d) represents the distance from the candidate centroid to (front, back, left, right, top, bottom), and l... mask Indicates a category mask.

[0036] The model learns the parameters of the regression branch by optimizing three loss functions: offset, size, and label. (shift，size) The smooth L1 loss, L2 loss, predicted angle, and label are interrelated. angle The bin-based loss L3 and the eight corner points and labels corner The corner loss L4 between them is calculated using the following formulas:

[0037]

[0038]

[0039]

[0040] Where p represents the predicted value.

[0041] The positioning accuracy prediction branch is achieved by calculating the Smooth L1 between the predicted value and the true HIoU value, as shown in the formula:

[0042]

[0043] Where, N p u represents the number of positive candidate points. p This represents a positive candidate point with a centrality label greater than 0.

[0044] The overall loss function is:

[0045]

[0046] Among them, L CG The regression loss of the CG layer is represented by the Smooth L1 loss function.

[0047] Beneficial effects:

[0048] This invention addresses the mismatch between classification scores and localization accuracy in single-stage object detectors and the excessive false detections found after analyzing the detection results of the DA-3DSSD model. It improves upon DA-3DSSD by adding a localization accuracy prediction branch and utilizing the predicted localization accuracy to improve the post-processing algorithm. Since there is no standard definition for localization accuracy labels, different localization accuracy metrics will yield different results. This invention proposes using GIoU as a localization accuracy metric. Furthermore, to improve the selection of the post-processing algorithm, a difficulty-adaptive metric HIoU is proposed, which divides the dataset into difficulty levels: HIoU equals IoU for simple targets, and HIoU equals GIoU for complex targets. Using the score obtained by multiplying HIoU by the classification score as the scoring basis in post-processing results in fewer false detections and more accurate target boxes. Experimental comparisons on the KITTI dataset show that the improved algorithm improves AP on the dataset, and its performance is highly competitive among recent methods. Attached Figure Description

[0049] Figure 1 The overall structure of a 3D target detection network with a positioning accuracy prediction branch;

[0050] Figure 2 This is a diagram of the backbone network structure for feature extraction.

[0051] Figure 3 This is a diagram of the network structure of the detection head;

[0052] Figure 4 A chart showing the assignment of difficulty level labels;

[0053] Figure 5 This is a schematic diagram of a two-dimensional minimum bounding box;

[0054] Figure 6 This is a schematic diagram of the three-dimensional minimum bounding box;

[0055] Figure 7 The flowchart is for the NMS post-processing algorithm;

[0056] Figure 8The image shows a comparison of the detection results of the present invention model and the DA-3DSSD model. The left column shows the detection results of the present invention, and the right column shows the detection results of DA-3DSSD. The black box is the truth box, the gray box is the detection box, and the suppressed false detections are marked with circles in the image. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] A three-dimensional target detection method with positioning accuracy prediction, the overall structure diagram is as follows: Figure 1 As shown. Specifically, it includes the following steps:

[0059] Step 1: Preprocess the input raw radar point cloud.

[0060] The specific preprocessing is as follows: the point cloud is cropped according to the range of the two-dimensional image, retaining the label of the current category while ignoring the labels of other categories; the target point cloud of the current category in each point cloud is saved separately for subsequent data augmentation; simple (0) or complex (1) labels are assigned to the target according to the occlusion, truncation, height and other conditions of the target, for training of the prediction branch of the difficulty adaptive localization accuracy label HIoU.

[0061] Step 2: Input the preprocessed point cloud into the constructed point cloud feature extraction backbone network to obtain the predicted candidate target center points and corresponding feature vectors. The main steps include the SA layer to gather local features and the CG layer to obtain the candidate center point positions and features.

[0062] The point cloud is input into the point feature extraction backbone network, and local features are aggregated step by step through three SA layers, such as... Figure 2 The Backbone section is shown. Each SA layer first samples the points, then groups the sampled points and performs feature aggregation to extract features around the sampled points. The sampling method combines density-aware point sampling (DPS) and feature-farthest point sampling (F-FPS) to retain more internal points for sparse targets.

[0063] The features extracted after three SA layers are used in the subsequent CG layer to predict the offset between the sampled points obtained from F-FPS and the instance center points. Subsequently, offsets are added to the sampled points to obtain candidate points. These candidate points are considered center points, and features are aggregated from the surrounding points of the last SA layer to calculate the center point density. Simultaneously, weights learned from the center point density using a multilayer perceptron (MLP) are used to reweight the features, resulting in the final candidate center points and features, as shown below. Figure 2 The CG layer section is shown.

[0064] Step 3: Construct a multi-task branch detection head. For each candidate target center point, predict the classification score, bounding box, and localization accuracy through classification, localization, and localization accuracy prediction branches. The detailed network structure of the detection head is as follows: Figure 3 As shown.

[0065] The specific structure of the classification branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes, and then through another MLP with an output node of 1, representing the predicted score of the target belonging to the given category (here referring to Car). Since the probability is a value between 0 and 1, the predicted value is then mapped to the range (0,1) by the Sigmoid function, thus obtaining the final classification confidence.

[0066] The specific structure of the localization branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes, then through an MLP with 30 output nodes, where 6 nodes output offset and size, 12 nodes output angle category, and 12 nodes output angle offset. Because the positional differences between targets are too large, the labels used in regression are encoded from the true labels to ensure network convergence. The predicted values ​​of the regression branch also need to be decoded to obtain the true positional information. The predicted values ​​include offset and related information for length, width, and height. The decoding formula is:

[0067]

[0068]

[0069] Among them, (x base ,y base ,z base The coordinates of the candidate center point are shown in the diagram. Orientation estimation uses a bin-based angle regression algorithm. The angle is divided into n intervals (n = 12 in this network), and the probability of the target angle belonging to a certain interval and the offset based on that interval are predicted. The network's prediction result (θ) cls ,θ res The actual target angle is obtained by decoding using the following formula:

[0070]

[0071] The specific structure of the localization branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes in the localization accuracy prediction branch, and then through an MLP with an output node of 1, used to predict the localization accuracy of the target box predicted by the localization branch. To achieve prediction, the location information and labels obtained from the localization branch are first converted into bounding boxes represented by 8 vertex coordinates. Then, by comparing the IoU, GIoU, and the proposed HIoU—three localization accuracy metrics—between the predicted box and the ground truth box, HIoU is selected as the target ground truth for localization accuracy. Finally, under the supervision of the HIoU ground truth, difficulty-adaptive localization accuracy prediction is achieved. Because the prediction results of the regression branch are involved in the calculation of the ground truth, it is important to prevent gradient backpropagation to the regression branch during training. The specific formula for calculating HIoU is:

[0072]

[0073]

[0074]

[0075] Here, "label" refers to the difficulty level label assigned during the preprocessing stage, and its assignment method is as follows: Figure 4 As shown.

[0076] The minimum bounding box volume representing the two 3D target boxes, such as Figure 5 As shown, the calculation formula is:

[0077]

[0078] Where y1 and y2 represent the y-axis coordinates of the top surface of bounding boxes A and B, respectively, and y3 and y4 represent the y-axis coordinates of the bottom surface of bounding boxes A and B, respectively.

[0079] The minimum bounding box for two-dimensional structures without rotation angle is calculated as follows: Figure 6 As shown, the minimum bounding box of two-dimensional bounding boxes A(x1,y1,x2,y2) and B(x3,y3,x4,y4) is C(min(x1,x3),min(y1,y3),max(x2,x4),max(y2,y4)), which is the box formed by the leftmost, topmost, rightmost, and bottommost bounding boxes of the two bounding boxes.

[0080] The method for calculating the minimum bounding box in two dimensions with a rotation angle is as follows:

[0081] First, rotate the coordinate axes to be parallel to the two sides of box A, by the rotation angle being either the predicted angle of the bounding box or the angle θ given in the label. In implementation, we choose an equivalent approach, rotating the two bounding boxes by opposite angles, as shown in the following formula, where R is the rotation matrix.

[0082] A R =R×A T B R =R×B T

[0083]

[0084] Then use Figure 6 The bounding box C1 in this case is obtained by the two-dimensional calculation method.

[0085] Similarly, rotate the coordinate axes to be parallel to the two sides of box B to obtain the bounding box C2 in this case.

[0086] The final minimum bounding box is the one with the smaller area between C1 and C2, i.e.

[0087] Finally, under the supervision of the HIoU ground truth, we achieve difficulty-adaptive localization accuracy prediction. Because the prediction results of the regression branch are involved in the calculation of the ground truth, we must be careful to prevent gradient backpropagation to the regression branch during training.

[0088] Step 4: Post-processing for filtering candidate boxes. The flowchart of the algorithm is as follows. Figure 7 As shown. The specific steps are as follows: The candidate bounding boxes and the scores obtained by multiplying the candidate bounding box classification scores by the predicted HIoU values ​​are used as inputs to the post-processing algorithm. The non-maximum suppression algorithm sorts the scores of the input boxes, continuously selects the boxes with the highest scores, and suppresses boxes that overlap too much with them, ultimately obtaining a series of non-overlapping boxes in 3D. Finally, candidate boxes with scores below a set threshold are removed from the detection result set.

[0089] Step 5: Construct a loss function to optimize network parameters during training.

[0090] The classification branch is learned under the supervision of continuous center-ness labels. The model learns network parameters by minimizing the cross-entropy loss between the predicted confidence score and the center-ness label, as shown in the formula:

[0091] L1(s,l ctrness )=-(l ctrness log s+(1-l ctrness log(1-s))

[0092]

[0093] Where s represents the predicted score of the candidate centroid. The second formula is used to calculate the centroid label of the candidate centroid, where (f,b,l,r,t,d) represents the distance from the candidate centroid to (front, back, left, right, top, bottom), and l... mask Indicates a category mask.

[0094] The model learns the parameters of the regression branch by optimizing three loss functions: offset, size, and label. (shift，size) The smooth L1 loss, L2 loss, predicted angle, and label are interrelated. angle The bin-based loss L3 and the eight corner points and labels corner The corner loss L4 between them is calculated using the following formulas:

[0095]

[0096]

[0097]

[0098] Where p represents the predicted value.

[0099] The positioning accuracy prediction branch is achieved by calculating the Smooth L1 between the predicted value and the true HIoU value, as shown in the formula:

[0100]

[0101] Where, N p u represents the number of positive candidate points. p This represents a positive candidate point with a centrality label greater than 0.

[0102] The overall loss function is:

[0103]

[0104] Among them, L CG The regression loss of the CG layer is represented by the Smooth L1 loss function.

[0105] The technical effects of the present invention will be further described below with reference to simulation experiments.

[0106] The simulation experiment of this invention uses the present invention and an existing 3D object detection network, and conducts experiments on the validation set of the KITTI dataset to obtain qualitative result images of point cloud completion, such as... Figure 8 As shown.

[0107] The existing technologies used in the simulation experiments include:

[0108] ①Chen X,Ma H,Wan J,et al.Multi-view 3D Object Detection Network forAutonomous Driving[C].2017IEEE Conference on Computer Vision and PatternRecognition(CVPR).Honolulu:IEEE,2017.6526-6534.

[0109] ②Zhou Y and Tuzel O.VoxelNet:End-to-End Learning for Point CloudBased 3D Object Detection[C].2018IEEE / CVF Conference on Computer Vision andPattern Recognition(CVPR).Salt Lake City:IEEE,2018.4490-4499.

[0110] ③Yan Y,Mao Y,and Li B.Second:Sparsely embedded convolutionaldetection[J].Sensors,2018,18(10):3337.

[0111] ④Shi S,Wang X,and Li H.PointRCNN:3D Object Proposal Generation andDetection from Point Cloud[C].2019IEEE / CVF Conference on Computer Vision andPattern Recognition(CVPR).Long Beach:IEEE,2019.770-779.

[0112] ⑤Yang Z,Sun Y,L Siu,et al.STD:Sparse-to-Dense 3D Object Detector forPoint Cloud[C].2019IEEE / CVF International Conference on Computer Vision(ICCV).Seoul:IEEE,2019.1951-1960.

[0113] ⑥Yang Z, Sun Y, Liu S, et al. 3DSSD: Point-Based 3D Single Stage ObjectDetector[C]. 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle: IEEE, 2020.11037-11045.

[0114] ⑦Zheng W, Tang W, Chen S, et al.CIA-SSD:Confident IoU-Aware Single-Stage Object Detector From Point Cloud[DB / OL].http: / / arxiv.org / abs / 2012.03015,2020.

[0115] ⑧He C, Zeng H, Huang J, et al.Structure Aware Single-Stage 3D ObjectDetection From Point Cloud[C].2020IEEE / CVF Conference on Computer Vision andPattern Recognition(CVPR),Seattle:IEEE,2020.11870-11879.

[0116] ⑨Ning J, Da F and Gai S. Density Aware 3D Object Single Stage Detector[J]. IEEE Sensors Journal, 2021, 21(20):23108-23117.

[0117] The following is combined with Figure 8 The simulation diagrams further illustrate the effects of the present invention. Figure 8 The left column shows the detection results of this invention, and the right column shows the detection results of DA-3DSSD. The black box represents the truth box, and the gray box represents the detection box. Suppressed false detections are marked with circles in the figure. Figure 8 It can be seen that the number of false detections in the model detection results of this invention is significantly less than that of DA-3DSSD, which demonstrates that the proposed method has a significant effect in suppressing false detections. Furthermore, from... Figure 8 By magnifying the target, it can be observed that the positioning accuracy prediction branch added to the model in this chapter makes the target positioning more accurate.

[0118] Using the average precision (AP) as the evaluation criterion for target detection results, the detection results of the present invention and the prior art are evaluated and plotted in Table 1.

[0119] Table 1. Quantitative test results of the present invention and prior art in the simulation experiment.

[0120]

[0121] As can be seen from Table 1, the present invention achieved the highest AP at all three difficulty levels on the validation set, verifying the effectiveness of the proposed method.

[0122] It should be noted that the above embodiments are merely examples for clear illustration and are not intended to limit the implementation methods. It is neither necessary nor possible to exhaustively list all possible implementation methods. All components not explicitly stated in this embodiment can be implemented using existing technology. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A three-dimensional target detection method with positioning accuracy prediction, characterized in that: Includes the following steps: S1. Preprocess the input raw radar point cloud; the specific method is: crop the point cloud according to the range of the two-dimensional image, retain the label of the current category and ignore the labels of other categories; save the target point cloud of the current category in each point cloud separately for subsequent data augmentation; Based on the target's occlusion, truncation, height, etc., simple or complex labels are assigned to the target. The simple label is "0" and the complex label is "1". This is used for training the prediction branch of the difficulty-adaptive localization accuracy label HIoU. S2. Input the preprocessed point cloud into the constructed point cloud feature extraction backbone network to obtain the predicted candidate target center points and corresponding feature vectors. S3. Construct a multi-task branch detection head, and predict the classification score, bounding box and localization accuracy for each candidate target center point through classification, localization and localization accuracy prediction branches; The positioning accuracy prediction branch uses HIoU as the target true value for positioning accuracy. The specific formula for calculating HIoU is as follows: Here, "label" refers to the difficulty level label assigned during the preprocessing stage; S4. The candidate bounding boxes are filtered through a post-processing algorithm to obtain the final prediction result; S5. Construct a loss function to optimize and train the network parameters.

2. The three-dimensional target detection method with positioning accuracy prediction according to claim 1, characterized in that, The main steps of the feature extraction backbone network for the preprocessed point cloud described in step S2 include the SA layer to gather local features and the CG layer to obtain the candidate center point positions and features.

3. The three-dimensional target detection method with positioning accuracy prediction according to claim 2, characterized in that: The specific method for aggregating local features in the SA layer is as follows: the point cloud is input into the point feature extraction backbone network, and local features are aggregated step by step through 3 SA layers; each SA layer first samples the points, then groups the sampled points and performs feature aggregation to extract the features around the sampled points; the sampling method combines density-sensitive point sampling and feature-based point sampling to retain more internal points for sparse targets; The specific method for obtaining candidate center point positions and features in the CG layer is as follows: the features extracted after three SA layers are used to predict the offset between the sampling points and the instance center points by the F-FPS obtained by the next CG layer; then, the offset is added to the sampling points to obtain candidate points; these candidate points are regarded as center points, and features are gathered from the surrounding points of the last SA layer and the center point density is calculated; at the same time, the features are reweighted using the weights learned from the center point density by the multilayer perceptron to obtain the final candidate center points and features.

4. The three-dimensional target detection method with positioning accuracy prediction according to claim 1, characterized in that: The multi-task branch detection head described in step S3 consists of a classification branch, a localization branch, and a localization accuracy prediction branch. The classification branch is responsible for predicting the target classification score, the localization branch is responsible for predicting the target bounding box, and the localization accuracy prediction branch is responsible for predicting the localization accuracy of the target bounding box predicted by the localization branch.

5. A three-dimensional target detection method with positioning accuracy prediction according to claim 4, characterized in that: The specific structure of the classification branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes in the classification branch, and then through an MLP with an output node of 1, which represents the score of the predicted target belonging to the given category; since the probability is a value between 0 and 1, after obtaining the predicted value, the score is mapped to the range of (0,1) by the Sigmoid function, thereby obtaining the final classification confidence.

6. The three-dimensional target detection method with positioning accuracy prediction according to claim 4, characterized in that: The specific structure of the localization branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes, then through an MLP with 30 output nodes, where 6 nodes output offset and size, 12 nodes output angle category, and 12 nodes output angle offset. Because the positional differences between targets are too large, the labels used in regression are encoded from the true labels to ensure network convergence. The predicted values ​​of the regression branch also need to be decoded to obtain the true positional information. The predicted values ​​include offset and related information for length, width, and height. The decoding formula is: Among them, (x base ,y base ,z base (θ) represents the coordinates of the candidate center point; direction estimation uses a bin-based angle regression algorithm; the angle is divided into n intervals, n is 12 in this network, and the probability of the target angle belonging to a certain interval and the offset based on the corresponding interval are predicted; the network's prediction result (θ) cls ,θ res The actual target angle is obtained by decoding using the following formula:

7. A three-dimensional target detection method with positioning accuracy prediction according to claim 4, characterized in that: The specific structure of the localization accuracy prediction branch is as follows: the features extracted by the backbone network first pass through a multilayer perceptron (MLP) consisting of 128 nodes, and then through an MLP with an output node of 1, to predict the localization accuracy of the target box predicted by the localization branch. To achieve prediction, the location information and labels obtained by the localization branch are first converted into bounding boxes represented by the coordinates of 8 vertices. Then, by comparing the IoU, GIoU, and the proposed HIoU, three localization accuracy metrics between the predicted box and the ground truth box, HIoU is selected as the target ground truth for localization accuracy. Finally, under the supervision of the HIoU ground truth, the prediction of localization accuracy with adaptive difficulty is achieved. Because the prediction results of the regression branch are involved in the calculation of the ground truth, it is important to prevent gradient backpropagation to the regression branch during training.

8. A three-dimensional target detection method with positioning accuracy prediction according to claim 1, characterized in that, The specific method of the post-processing algorithm for filtering candidate bounding boxes in step S4 is as follows: the candidate box and the score obtained by multiplying the candidate box classification score by the predicted value of HIoU are used as the input of the post-processing algorithm. The non-maximum suppression algorithm sorts the scores of the input boxes, continuously selects the box with the highest score and suppresses the boxes that overlap with it too much, and finally obtains a series of boxes that do not overlap in three dimensions. Finally, the candidate boxes with scores lower than the set threshold are removed from the detection result set.

Images