Target detection method, device and equipment

Abstract

The invention provides a target detection method, device and equipment. The method provided by the embodiment of the invention comprises the steps of obtaining point cloud data collected by a laser radar; inputting the point cloud data into a plurality of multi-layer sensors in a coding module to obtain coding features of the point cloud data; inputting the coding features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; and obtaining a target detection result according to the sparse features of the point cloud data. According to the method of the embodiment, the efficiency and accuracy of target detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a target detection method, device and equipment. Background Art

[0002] The point cloud data obtained by LiDAR is sparsely and irregularly distributed in continuous space. Most existing object detectors use grid-based encoding methods to convert point clouds into regular grid structures, such as columns, voxels, range view encoding, etc.

[0003] Among the related technologies, column-based coding can be directly compatible with two-dimensional image processing technology and is more widely used in industry. However, in the point cloud coding process, the features extracted only by maximum pooling often lack fine-grained information, which limits subsequent processing and leads to low accuracy of detection results. Summary of the invention

[0004] The present invention provides a target detection method, device and equipment, which solve the problem in the related art that only extracting features through maximum pooling leads to loss of small target features and change of relative position in the original area, thereby improving the efficiency and accuracy of target detection.

[0005] The present invention provides a target detection method, comprising the following steps.

[0006] Get point cloud data collected by LiDAR; Inputting the point cloud data into a plurality of multi-layer perceptrons in an encoding module to obtain encoding features of the point cloud data; Inputting the encoded features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; According to the sparse features of the point cloud data, a target detection result is obtained.

[0007] According to a target detection method provided by the present invention, the encoding module includes: Two multilayer perceptrons and pooling modules in series.

[0008] According to a target detection method provided by the present invention, the sparse convolutional network module includes: Sparse convolution module and attention module; among them,

[0009] The sparse convolution module includes a sparse convolution layer and a submanifold convolution layer; The attention module includes channel and spatial dual attention modules.

[0010] According to a target detection method provided by the present invention, obtaining a target detection result according to the sparse features of the point cloud data includes: According to the sparse features of the point cloud data, global features and local features of the point cloud data are obtained; According to the global features and the local features, a target detection result is obtained.

[0011] According to a target detection method provided by the present invention, obtaining global features and local features of point cloud data according to the sparse features of the point cloud data includes: Downsampling the sparse features to obtain global features of the point cloud data; Some of the sparse features are downsampled to obtain local features of the point cloud data.

[0012] According to a target detection method provided by the present invention, obtaining a target detection result according to the global features and the local features includes: The global features and the local features are concatenated and fused to obtain a fusion result of the global features and the local features; The fusion result of the global feature and the local feature is input into a preset detection head to obtain the target detection result.

[0013] The present invention also provides a target detection device, comprising the following modules: An acquisition module is used to obtain point cloud data collected by the laser radar; A processing module, used for inputting the point cloud data into a plurality of multi-layer perceptrons in the encoding module to obtain encoding features of the point cloud data; Inputting the encoded features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; The detection module is used to obtain target detection results based on the sparse features of the point cloud data.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the target detection methods described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the target detection method as described above is implemented.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the target detection methods described above.

[0017] The target detection method, device and equipment provided by the present invention obtain the coding features of the point cloud data by inputting the point cloud data into multiple multi-layer perceptrons in the coding module; wherein the coding module includes multiple multi-layer perceptrons and pooling modules connected in series, thereby improving the recognition ability of the details of the point cloud data through multi-form aggregation, effectively obtaining fine-grained point cloud coding features, solving the problem in the related art that only the maximum pooling is used to extract features, resulting in the loss of small target features and the change of relative position in the original area, thereby improving the efficiency and accuracy of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a flow chart of the target detection method provided by the present invention.

[0020] Figure 2 It is a schematic diagram of the encoding module provided by the present invention.

[0021] 3a-3f are comparison diagrams of experimental results of target detection provided by the present invention.

[0022] Figure 4 It is a structural schematic diagram of the target detection device provided by the present invention.

[0023] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] Combine the following Figure 1-Figure 5 The object detection method, device and apparatus of the present invention are described.

[0026] In order to facilitate a clearer understanding of the technical solutions of the embodiments of the present application, some technical contents related to the embodiments of the present application are first introduced.

[0027] Computer vision plays an important role in deep learning, and 3D perception is a key challenge in this field. In recent years, 3D perception technology has made significant progress in the fields of artificial intelligence and autonomous driving. Using sensors such as lidar and cameras to obtain spatial data and combining it with deep learning and machine learning algorithms, it is possible to identify, locate and track objects in 3D space. These technologies enhance the environmental perception capabilities of autonomous driving systems, which is conducive to safe and efficient autonomous driving. In autonomous driving systems, point cloud object detection is key. Point cloud data obtained by sensors such as lidar is combined with deep learning algorithms to detect and identify vehicles, pedestrians, obstacles and other surrounding objects. The 3D accuracy of this data can more accurately capture and understand the location and shape of objects, providing reliable and precise perception for autonomous driving systems.

[0028] The point cloud data obtained by LiDAR is sparsely and irregularly distributed in continuous space. Most existing object detectors use grid-based encoding methods to convert point clouds into regular grid structures, such as column, voxel, and range view encoding, making the application of regular operators possible. Researchers generally believe that voxel-based encoding has higher detection accuracy than column-based encoding because it performs additional quantization on the height of the point cloud, thereby reducing information loss. Compared with voxel-based encoding, column-based encoding can be directly compatible with 2D image processing technology and is more widely used in industry. Voxel-based encoding divides the input point cloud into a regular 3D voxel grid and uses sparse 3D convolution to learn multi-scale geometric information. After encoding, standard 2D convolution fuses multi-scale feature information and then passes it to the detection head. Column-based encoding performs maximum pooling in the vertical direction of the input point cloud to extract the initial features of each column. These features are then projected into a pseudo 2D image of a bird's-eye view (BEV) for multi-scale feature extraction. Finally, detection is performed using a method similar to voxel encoding.

[0029] In recent studies, methods for extracting features from encoded point cloud data to the backbone network include sparse 3D convolution, sparse 2D convolution, and regular 2D convolution. Each method has its own unique advantages and limitations and is suitable for different application scenarios. The traditional 2D convolution method is to map point cloud data to a 2D image, input the point cloud features into the sparse convolution backbone network, and use existing image processing techniques for feature extraction. This method can be seamlessly integrated into existing deep learning frameworks and is easy to implement. However, the 3D to 2D mapping may lead to the loss of important spatial information, which requires a carefully designed process to retain as much useful information as possible. Sparse 3D convolution directly operates on point cloud data in 3D space and makes full use of the geometric information of the point cloud. Its sparsity provides high computational efficiency for large-scale point clouds and avoids intensive computation of the entire space. However, implementing sparse convolution is relatively complex and requires special data structures such as octrees or kd trees for efficient indexing and searching, which increases the complexity of the algorithm. After projecting the point cloud data onto a 2D plane, sparse 2D convolution can extract features using established 2D convolution techniques, simplifying the implementation of the algorithm. However, this may result in the loss of some 3D spatial information, especially during the conversion process. In outdoor object detection, considering the lack of spatial overlap between vehicles and pedestrians, it is feasible to use sparse 2D convolution to reduce the amount of training computation.

[0030] Previous methods have made important contributions to point cloud object detection. However, we found that previous methods generally have a high false positive rate. We believe that the reason is that in the point cloud encoding process, the features extracted only by max pooling often lack fine-grained information, which limits subsequent processing. Similar to sparse point cloud features, feature maps may contain redundant information during the convolution process, which will interfere with the final prediction results and reduce detection performance.

[0031] Figure 1 is one of the flow charts of the target detection method provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 101: Obtain point cloud data collected by the laser radar.

[0032] Specifically, in the point cloud encoding process, features extracted only by maximum pooling often lack fine-grained information, which limits subsequent processing and leads to low accuracy of detection results.

[0033] In order to solve the above problem, in the embodiment of the present application, the point cloud data collected by the laser radar is first obtained. Optionally, the point cloud data collected by the laser radar can be the point cloud data of target objects such as pedestrians, vehicles, buildings, etc. around the driving vehicle equipped with the laser radar.

[0034] Step 102: input the point cloud data into a plurality of multi-layer perceptrons in the encoding module to obtain encoding features of the point cloud data.

[0035] Specifically, after acquiring the point cloud data collected by the laser radar, in the embodiment of the present application, the point cloud data is input into multiple multi-layer perceptrons in the encoding module, so as to quickly and accurately obtain the encoding features of the point cloud data. Optionally, the encoding module includes multiple multi-layer perceptrons and pooling modules connected in series, thereby improving the recognition ability of the details of the point cloud data through multi-form aggregation, effectively obtaining fine-grained point cloud encoding features, solving the problem of small target feature loss and relative position change in the original area caused by extracting features only through maximum pooling in related technologies, and improving the efficiency and accuracy of target detection.

[0036] Step 103: Input the encoded features of the point cloud data into the sparse convolutional network module to obtain the sparse features of the point cloud data.

[0037] Specifically, after inputting the point cloud data into the multiple multi-layer perceptrons in the encoding module to obtain the encoded features of the point cloud data, the embodiment of the present application further inputs the encoded features of the point cloud data into the sparse convolutional network module to obtain the sparse features of the point cloud data. Optionally, the sparse convolutional network module can be based on sparse three-dimensional convolution, sparse two-dimensional convolution or conventional two-dimensional convolution modules, which is not specifically limited in the embodiment of the present application. This method can be seamlessly integrated into the existing deep learning framework and is easy to implement.

[0038] Step 104: Obtain target detection results based on the sparse features of the point cloud data.

[0039] Specifically, after the encoded features of the point cloud data are input into the sparse convolutional network module to obtain the sparse features of the point cloud data, the sparse features of the point cloud data can be input into a preset detection head to obtain the target detection results.

[0040] The method of the above embodiment obtains the encoding features of the point cloud data by inputting the point cloud data into multiple multi-layer perceptrons in the encoding module; wherein the encoding module includes multiple multi-layer perceptrons and pooling modules connected in series, thereby improving the recognition ability of the details of the point cloud data through multi-form aggregation, effectively obtaining fine-grained point cloud encoding features, and solving the problem in the related art that only the maximum pooling is used to extract features, resulting in the loss of small target features and the change of relative position in the original area, thereby improving the efficiency and accuracy of target detection.

[0041] In one embodiment, the encoding module includes: Two multilayer perceptrons and pooling modules in series.

[0042] Specifically, in previous work, point cloud encoders can be divided into pure point cloud encoders, voxel encoders, and column encoders. The column encoder discretizes the point cloud into a structured grid and aggregates the points in each unit into a preliminary feature representation. That is, the point cloud is organized into vertical columns for maximum pooling. After analysis, it was found that fine-grained features are lost in existing point cloud encoders, which may lead to information loss, oversimplification, and even loss of small target features, thereby affecting the recognition and processing of these targets, as well as the preservation of relative position information in the original area.

[0043] Therefore, the encoding module in the present application includes multiple multi-layer perceptrons and pooling modules connected in series, thereby improving the network's ability to recognize the details of point cloud data through multi-form aggregation, effectively obtaining fine-grained point cloud coding features, and improving the efficiency and accuracy of target detection.

[0044] It should also be noted that, after repeated experiments and verifications, the inventors of the present application found that if the coding module only includes a single multi-layer perceptron MLP to extract grid features, there will still be problems with small target feature loss and relative position changes in the original area, while the presence of multiple multi-layer perceptron MLPs in the coding module will lead to high power consumption and the feature extraction results will decrease. Therefore, after repeated experiments and verifications, it was found that the coding module only includes two multi-layer perceptrons and a pooling module in series, which can not only more effectively obtain fine-grained point cloud coding features and improve the efficiency and accuracy of target detection, but also effectively balance power consumption.

[0045] For example, Figure 2 As shown, the encoding module in the embodiment of the present application is a multi-layer additive encoder (MAP) for fine-grained feature learning. The multi-layer additive encoder includes two multi-layer perceptrons connected in series, a max pooling module, and a MAF pooling module.

[0046] First, a multi-layer additive encoder (MAP) is used for feature extraction, and then the element summation of max pooling and MAF pooling is used to extract the coding features of the pillars, thereby obtaining fine-grained features for downstream tasks. That is, the point cloud features obtained by the two-layer MLP are respectively passed through max pooling and MAF pooling, which can more effectively obtain fine-grained point cloud coding features, improve the efficiency and accuracy of target detection, and effectively balance power consumption.

[0047] like Figure 2 As shown, the first multilayer perceptron MLP in the multilayer additive encoder is used for preliminary feature learning. Assuming the input feature is x, the mathematical expression of the MLP operation is:

[0048] The MLP layer consists of a linear layer, a batch normalization layer, and a ReLU activation function. Next, max pooling is applied to the learned features, which are then concatenated with the original features. After this, the features pass through another MLP layer, after which max pooling and MAF pooling are performed. These two sets of features are then element-wise summed to obtain the features of the column. In general, this process can be expressed as:

[0049] where p considers a scene where the i-th column contains a set of Ni LiDAR points, denoted by p∈RNi×7, where each point is characterized by its (x, y, z) coordinates and the difference of x, y, z relative to the center point of the voxel grid. The MLP layer consists of a linear layer, a batch normalization layer, and a ReLU activation function. Concat refers to channel-level connections. Max represents the operation of obtaining the maximum value along the batch dimension. Add represents element-wise summation.

[0050] By adding Max and MAF, we get the point cloud encoding features with fine-grained information. MAF can be expressed as:

[0051] The CSD consists of a layer-by-layer convolution with a convolution kernel, a Swish activation function, and a convolution. The CRD consists of a layer-by-layer convolution with a convolution kernel, a ReLU activation function, and a convolution. σ represents the Sigmoid function, Mul represents element-wise multiplication, and Sum represents batch dimension summation. We use the Ds formula to extract global features and the Dr formula to extract local features. Subsequently, multi-feature fusion is performed to obtain rich feature information with global perception. Experiments have shown that this module can capture fine-grained information during training, thereby enhancing the robustness of subsequent training stages.

[0052] In the method of the above embodiment, the encoding module includes two multi-layer perceptrons and a pooling module that are serially connected, which can not only more effectively obtain fine-grained point cloud coding features and improve the efficiency and accuracy of target detection, but also effectively balance power consumption.

[0053] In one embodiment, the sparse convolutional network module includes: Sparse convolution module and attention module; among them,

[0054] The sparse convolution module includes sparse convolution layer and submanifold convolution layer; The attention module includes channel and spatial dual attention modules.

[0055] Specifically, the sparse convolutional network module in the embodiment of the present application includes a sparse convolutional module and an attention module. The sparse convolutional module includes a sparse convolutional layer and a submanifold convolutional layer, which are used to perform multi-scale extraction of sparse convolutional features on point clouds.

[0056] Optionally, the sparse convolution module in the embodiment of the present application is composed of a sparse convolution layer and a submanifold layer. Submanifold convolution (SmConv) is used to perform channel transformation on voxel features to maintain the sparsity of the point cloud. Sparse convolution (SpConv) improves the convolution efficiency by only calculating the convolution of non-empty voxels. That is, the sparse submanifold convolution strictly operates on non-empty features, maintaining the original geometric shape. It can output features in empty positions, providing flexibility and expanding the receptive field during the convolution process. Therefore, each feature block is sparsely convolved to capture multi-scale features, and then channel transformation is performed through multiple submanifold convolution layers to enrich the nonlinear representation of the features. Considering that there are a large number of sparse points in point cloud data, using sparse convolution on dense convolution can significantly reduce computational overhead.

[0057] In addition, it should be noted that in the sparse CNN, the present application suppresses unimportant channels and enhances important channels through the attention mechanism of channel and spatial dimensions, and identifies important areas by paying extra attention to important areas, thereby improving the representation ability of feature maps; that is, the power of sparse convolutional blocks is combined with the double attention mechanism to capture rich spatial layers and channel dependencies from sparse features, thereby generating robust feature representations and significantly improving the network's discrimination and feature extraction capabilities. The details are as follows:

[0058] Where S represents the input sparse feature. This application introduces sparse feature extraction technology, uses SparseAvgPool and SparseMaxPool operations to perform average pooling and maximum pooling on the input feature S respectively, and extracts representative information of the sparse feature. In addition, this application uses SRS (the sparse feature learning module of SubMConv), combined with the ReLU activation function and the SubMConv operation, to integrate features through element-wise addition and element-wise multiplication. Specific operations are performed along the spatial dimension, as follows:

[0059] Where avg is the average operation of the channel dimension, max is the maximum operation of the channel dimension, σ represents the Sigmoid function, and Concat is the channel-level connection operation. This application uses the power of sparse convolution blocks in combination with dual attention mechanisms. These mechanisms are able to skillfully identify significant channels and dynamically adjust the weights and spatial positions of feature channels. This complex process is crucial to improving the overall performance of the network. That is, a sparse dual attention mechanism is used for the features obtained by downsampling to enhance useful information and suppress useless information. SparseConv and SubMConv are then used for subsequent feature learning. By strategically combining these operations, the method of the present application can capture rich spatial layers and channel dependencies from sparse features, generate robust feature representations, and thus significantly improve the network's discrimination and feature extraction capabilities.

[0060] The method of the above embodiment combines the power of sparse convolution blocks with the dual attention mechanism, captures rich spatial layers and channel dependencies from sparse features, suppresses unimportant channels and enhances important channels through the attention mechanism of channel and spatial dimensions, and identifies important areas by paying extra attention to important areas, thereby improving the representation ability of feature maps, generating robust feature representations, and significantly improving the network's discrimination and feature extraction capabilities, thereby effectively improving the efficiency and accuracy of target detection.

[0061] In one embodiment, the target detection result is obtained according to the sparse features of the point cloud data, including: According to the sparse features of the point cloud data, the global features and local features of the point cloud data are obtained; According to the global features and local features, the target detection results are obtained.

[0062] Specifically, in the task of 3D point cloud object detection, expanding the receptive field is crucial for accurately identifying objects of different sizes and shapes. CNN has achieved great success in the field of image processing due to a deep understanding of the limitations of existing CNN architectures in processing point cloud data, but its fixed receptive field is often insufficient to capture long-range feature associations in point cloud data. That is, although CNN performs well in processing local features, its fixed receptive field limits the capture of global contextual information.

[0063] In order to address this limitation, this application proposes a simple and effective neck design. After obtaining the sparse features of the point cloud data, the global features and local features of the point cloud data are further obtained based on the sparse features of the point cloud data. By integrating local and global feature learning, the receptive field of the model can be significantly improved, thereby effectively improving the efficiency and accuracy of target detection.

[0064] In one embodiment, the global features and local features of the point cloud data are obtained according to the sparse features of the point cloud data, including: Downsample the sparse features to obtain the global features of the point cloud data; Some features in the sparse features are downsampled to obtain local features of the point cloud data.

[0065] Specifically, a local and global fusion module (CTGC) is designed in the embodiment of the present application. This module uses two branches to obtain local features and global features respectively. In branch one, the traditional CNN architecture is used to process the point cloud features that have been downsampled 8 times. This branch takes advantage of the advantages of CNN in local feature extraction and provides the model with rich local geometric information. Through multi-layer convolution and pooling operations, the spatial resolution of the features is gradually reduced, while the number of channels is increased to extract higher-level abstract features. In branch two, in order to achieve more comprehensive context information capture, the present application introduces SwinTransformer to extract global features, which limits the self-attention calculation to non-overlapping local windows while allowing cross-window connections to improve efficiency. This hierarchical architecture provides a flexible modeling scale and effectively extracts global context information. The long-distance dependencies in the point cloud data are captured through a global self-attention mechanism, thereby realizing the learning of global features. In this application, local features and global features are fused to expand the receptive field, and a splicing operation is used to combine the outputs of the two branches.

[0066] Optionally, the present application combines the self-attention mechanism of convolutional fusion to capture local details and global context, combining local feature accuracy with global semantic information, that is, splicing the local features of branch one and the global features of branch two along the channel dimension to form a feature representation that integrates local details and global context. This fusion strategy not only retains the fineness of local features, but also incorporates the semantic information of global features, providing richer feature support for subsequent target detection tasks. This hybrid architecture not only takes advantage of CNN's advantages in local feature extraction, but also captures global context information. The fusion strategy is simple and effective, and significantly improves the model's ability to detect targets of different scales.

[0067] The method of the above embodiment splices local features with global features along the channel dimension. This fusion strategy not only retains the fineness of local features, but also incorporates the semantic information of global features, providing richer feature support for subsequent target detection tasks, and significantly improving the detection capability and accuracy of targets of different scales.

[0068] Exemplarily, as shown in Figures 3a-3f, the model of the present application is put together with the existing models SECOND method, PointPillars method, PointRCNN method and IA-SSD method to observe and compare the results. Figures 3(a), (b), (c) and (d) show the visualization results of the above methods. Figure 3(e) shows the ground truth officially provided by KITTI; Figure 3(f) shows the visual detection results of our proposed method. The orange box indicates the false detection box, and the red box indicates the undetected box. That is, the quantitative results on the KITTI validation set, the green, yellow and blue boxes are the predicted 3D boxes of cars, bicycles and pedestrians, the sky blue boxes are the 3D boxes of the ground truth, and the orange boxes are the unmarked areas of the ground truth. In the SECOND and PointPillars visualizations, many false detections are obvious. Similarly, the PointRCNN and IA-SSD visualizations show certain missed detections and false detections. In contrast, the visualization results of our method accurately identify the real instances with only a small number of false detections.

[0069] The method of the above embodiment first uses the encoding module to obtain fine-grained point cloud encoding features, and then uses the sparse convolutional network module to further enhance useful information and suppress useless information. Finally, the global and local context feature fusion module with 8 downsampling and 16 downsampling is sent to the detection head for target detection. Through a large number of experiments on the KITTI and nuScenes datasets, the method of the present application has a lower false detection rate than the existing detection method, which proves the versatility and effectiveness of the target detection method of the present application, and is superior to some mainstream detectors. The results show that the method of the present application has achieved significant performance improvements.

[0070] The target detection device provided by the present invention is described below. The target detection device described below and the target detection method described above can be referred to in correspondence with each other. Figure 4 As shown, including: An acquisition module 410 is used to acquire point cloud data collected by a laser radar; The processing module 420 is used to input the point cloud data into the multiple multi-layer perceptrons in the encoding module to obtain the encoding features of the point cloud data; Input the encoded features of the point cloud data into the sparse convolutional network module to obtain the sparse features of the point cloud data; The detection module 430 is used to obtain target detection results based on the sparse features of the point cloud data.

[0071] Optionally, the encoding module includes: Two multilayer perceptrons and pooling modules in series.

[0072] Optionally, a sparse convolutional network module, comprising: Sparse convolution module and attention module; among them,

[0073] The sparse convolution module includes sparse convolution layer and submanifold convolution layer; The attention module includes channel and spatial dual attention modules.

[0074] Optionally, the detection module 430 is used to obtain global features and local features of the point cloud data according to the sparse features of the point cloud data; According to the global features and local features, the target detection results are obtained.

[0075] Optionally, the detection module 430 is used to downsample the sparse features to obtain global features of the point cloud data; Some features in the sparse features are downsampled to obtain local features of the point cloud data.

[0076] Optionally, the detection module 430 is used to concatenate and fuse the global features and the local features to obtain a fusion result of the global features and the local features; The fusion results of global features and local features are input into the preset detection head to obtain the target detection results.

[0077] Figure 5 The schematic diagram of the physical structure of an electronic device is illustrated, and the electronic device may include: a processor 510, a communication interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communication interface 520 and the memory 530 communicate with each other through the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute the target detection method, and the method includes: obtaining point cloud data collected by the laser radar; inputting the point cloud data into multiple multi-layer perceptrons in the encoding module to obtain the encoding features of the point cloud data; inputting the encoding features of the point cloud data into the sparse convolutional network module to obtain the sparse features of the point cloud data; and obtaining the target detection result according to the sparse features of the point cloud data.

[0078] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0079] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the target detection method provided by the above methods, which includes: acquiring point cloud data collected by a lidar; inputting the point cloud data into multiple multi-layer perceptrons in an encoding module to obtain encoding features of the point cloud data; inputting the encoding features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; and obtaining target detection results based on the sparse features of the point cloud data.

[0080] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it is implemented to execute the target detection method provided by the above-mentioned methods, and the method includes: acquiring point cloud data collected by a laser radar; inputting the point cloud data into multiple multi-layer perceptrons in an encoding module to obtain encoding features of the point cloud data; inputting the encoding features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; and obtaining target detection results based on the sparse features of the point cloud data.

[0081] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0082] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target detection method, characterized in that: include: Get point cloud data collected by LiDAR; Inputting the point cloud data into a plurality of multi-layer perceptrons in an encoding module to obtain encoding features of the point cloud data; Inputting the encoded features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; According to the sparse features of the point cloud data, a target detection result is obtained.

2. The target detection method according to claim 1, characterized in that: The encoding module comprises: Two multilayer perceptrons and pooling modules in series.

3. The target detection method according to claim 1, characterized in that: The sparse convolutional network module comprises: Sparse convolution module and attention module; among them, The sparse convolution module includes a sparse convolution layer and a submanifold convolution layer; The attention module includes channel and spatial dual attention modules.

4. The target detection method according to claim 1, characterized in that: Obtaining a target detection result according to the sparse features of the point cloud data includes: According to the sparse features of the point cloud data, global features and local features of the point cloud data are obtained; According to the global features and the local features, a target detection result is obtained.

5. The target detection method according to claim 1, characterized in that: The step of obtaining the global features and local features of the point cloud data according to the sparse features of the point cloud data includes: Downsampling the sparse features to obtain global features of the point cloud data; Some of the sparse features are downsampled to obtain local features of the point cloud data.

6. The target detection method according to claim 4, characterized in that: Obtaining the target detection result according to the global features and the local features includes: The global features and the local features are concatenated and fused to obtain a fusion result of the global features and the local features; The fusion result of the global feature and the local feature is input into a preset detection head to obtain the target detection result.

7. A target detection device, characterized in that: include: An acquisition module is used to obtain point cloud data collected by the laser radar; A processing module, used for inputting the point cloud data into a plurality of multi-layer perceptrons in the encoding module to obtain encoding features of the point cloud data; Inputting the encoded features of the point cloud data into a sparse convolutional network module to obtain sparse features of the point cloud data; The detection module is used to obtain target detection results based on the sparse features of the point cloud data.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the target detection method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the target detection method according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the target detection method according to any one of claims 1 to 6 is implemented.

Images