Target object detection method and device, storage medium and self-driven equipment

Abstract

The invention provides a target object detection method and device, a storage medium and self-driven equipment, and relates to the field of mobile positioning. According to the target object detection method, the point cloud data are captured through the laser radar, the point cloud data are clustered, and the feature points are identified according to the clustered point cloud data, so that the target object is identified, and the technical problems that the feature points cannot be accurately identified or the feature points are lost in the point cloud data processing process in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mobile positioning, and in particular to a target object detection method and device, a storage medium and a self-driving device. BACKGROUND

[0002] With the rapid development of artificial intelligence devices, the positioning and navigation technology of robots has also developed rapidly. In the prior art, point clouds collected by a laser radar are analyzed to determine the category of the object corresponding to the point cloud data, thereby identifying a target object and achieving positioning.

[0003] However, in actual applications, the driving environment of the device and / or the working environment are complex, and the objects around the device and / or the distance between the objects are constantly changing, resulting in a large amount of point cloud data, and thus the phenomenon of feature point loss occurs in the process of processing the point cloud data, thereby reducing the recognition accuracy. SUMMARY

[0004] The main purpose of the present application is to provide a target object detection method and device, which aims to accurately detect target objects in a working scenario and improve the accuracy of device positioning to meet the needs of users.

[0005] In a first aspect, the present application provides a target object detection method, which comprises:

[0006] collecting N frames of point cloud data within a preset time period;

[0007] dividing the space occupied by the point cloud data into a plurality of voxels of a preset specification, and clustering the point cloud data within each voxel;

[0008] processing the voxels by a feature extraction algorithm and / or a pattern recognition algorithm to identify target objects in the target scenario.

[0009] In one example, the clustering of the point cloud data within each voxel in the method comprises:

[0010] dividing the space occupied by the point cloud data into a plurality of voxels of a preset specification, and clustering the point cloud data within each voxel;

[0011] performing Euclidean clustering on the point cloud points contained in each sub-voxel to classify the sub-voxels occupied by the point cloud points within a preset range in the space into one category as a voxel.

[0012] The target object detection method provided in this application captures point cloud data with LiDAR, performs clustering processing on the point cloud data, and then identifies feature points based on the clustered point cloud. This solves the problem of inaccurate feature point identification or feature point loss during point cloud data processing in the prior art, thereby effectively improving the positioning accuracy and positioning stability of the self-driven device.

[0013] A second aspect of this application provides a target object detection apparatus, the apparatus comprising:

[0014] The acquisition module collects N frames of point cloud data within a preset time period;

[0015] The processing module divides the point cloud data into several voxels of a preset size according to the space occupied by the point cloud data, and performs cluster processing on the point cloud data within each voxel.

[0016] The identification module processes the voxels using feature extraction algorithms and / or pattern recognition algorithms to identify target objects in the target scene.

[0017] A third aspect of the embodiments of this application provides a self-driving device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0018] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0019] The target object detection method provided in this application captures point cloud data with LiDAR, performs clustering processing on the point cloud data, and then identifies feature points based on the clustered point cloud. This solves the problem of inaccurate feature point identification or feature point loss during point cloud data processing in the prior art, thereby effectively improving the positioning accuracy and positioning stability of the self-driven device. Attached Figure Description

[0020] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings:

[0021] Figure 1 This is a schematic diagram illustrating an application scenario of the target object detection method provided in one embodiment of this application;

[0022] Figure 2 This is a flowchart of a target object detection method provided in one embodiment of this application;

[0023] Figure 3A This is point cloud data of a target object detection method provided in one embodiment of this application. Figure 1 ;

[0024] Figure 3B Point cloud data of the target object detection method provided by an embodiment of the present application Figure 2 ;

[0025] Figure 4 The structural schematic diagram of the target object detection device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0026] The main purpose of the embodiments of the present application is to provide a target object detection method and device, aiming to realize accurate detection of target objects in a working scene, improve the accuracy of device positioning, and meet the needs of users.

[0027] Laser radar scans and identifies objects The process of using a laser radar sensor to scan the surrounding environment and identifying the position, shape and size of surrounding objects (such as trees, flower beds, fences, etc.) by processing the scanned data. According to the result of identifying objects, the intelligent mower can finally realize autonomous navigation, obstacle avoidance and accurate mowing.

[0028] The current laser radar and its supporting algorithm are based on repeated scanning of laser point cloud, and different line bundles can process the repeated point cloud, such as line-surface feature extraction. However, for non-repetitive scanning laser radars, it is difficult to calculate line features and corner points. In actual slam applications, the processing method for non-repetitive 3D laser point cloud is relatively simple, which is only uniformly interval sampling and directly using a voxel filter for processing. Although it can also uniformly process data, it is easy to lose some relatively stable features in outdoor scenes.

[0029] In view of the deviation in the process of identifying object features, the embodiments of the present application provide a target object detection method, which scans the surrounding environment by laser radar, extracts and analyzes point cloud data, processes the data by using voxel clustering technology, removes noise and distortion, finally identifies the target object, and realizes accurate positioning, thereby improving the working efficiency and safety of the mower.

[0030] The method provided by the embodiments of the present application performs distance-based clustering operation on 3D point cloud, that is, the distance of all point clouds is judged according to a certain distance threshold, and the point clouds with similar distances are classified and processed. In the meantime, the point clouds after clustering are removed from noise and uniformly processed, which improves the accuracy of target object identification, thereby improving the stability of real-time positioning of the machine in outdoor movement.

[0031] The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript.

[0032] In order to make the technical solutions and advantages in the embodiments of the present application clearer, the exemplary embodiments of the present application are further described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and are not exhaustive of all embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0033] The application environment of the target object detection method provided by the embodiments of the present application is briefly described as follows:

[0034] Referring to Figure 1 The object detection system 10 provided by the embodiments of the present application comprises at least a processor 101 and a sensor 102, the processor 101 is connected with the sensor 102, the processor acquires point cloud data collected by the sensor, wherein the sensor 102 can comprise a radar laser ranging sensor and the like.

[0035] The processor 101 acquires the point cloud data collected by the sensor 102, divides a space occupied by the point cloud data into a plurality of voxels of a preset specification, and performs clustering processing inside each voxel, performs mean value processing and distortion correction processing on the point cloud data, and finally processes the voxels by a feature extraction algorithm and / or a pattern recognition algorithm to identify a target object in a target scene.

[0036] Referring to Figure 2 The positioning processing method provided by the embodiments of the present application is applied to a processor 201, and the method provided by the embodiments of the present application is applied to identification of a target object in device operation. The target object detection method provided by the embodiments of the present application comprises the following steps 201-203.

[0037] Step 201, collect N frames of point cloud data in a preset time period.

[0038] The self-driving device enters a scanning area, a laser radar scans a target calibration object, obtains laser data, extracts laser point cloud from the scanned laser data, calculates the number n of points on the calibration object, collects the driving position of the device, collects adjacent two frames of point cloud data before and after, and performs matching processing.

[0039] The target calibration object can comprise a pre-set light board, can be an object in a preset area, or can be an obstacle in a driving path.

[0040] The laser data collected initially is first judged as reliable data before processing, for example, whether the driving state of the vehicle is in stable driving, road condition information, and radar point number are comprehensively considered to judge whether the obtained radar data is reliable.

[0041] After acquiring multiple sets of continuous frame laser data, the laser data is clustered and filtered to remove noise and outliers, and the laser data is projected according to the pitch angle and roll angle in the device operating environment to ensure the reliability of the laser data.

[0042] The self-driving device includes a device configured with a motor and a processor, which can send signals to control the motor to drive the device to move, such as a smart robot, a smart mower, or other working devices.

[0043] The target object can include a specific plant (such as weeds on the lawn or ornamental plants that need to be preserved), an obstacle (such as a stone, a junk, a fence, etc.), a charging pile, a user, an animal, and other objects that can indicate a specific target.

[0044] The 3D SLAM or 2D SLAM technology based on laser point cloud uses a multi-dimensional laser sensor to obtain multi-dimensional spatial point data, such as two-dimensional or three-dimensional spatial point data, and then performs pose estimation through scan matching between adjacent point cloud frames and establishes a complete point cloud map.

[0045] The above-mentioned preset time period is set by customer demand, and specifically can be 1 second, 0.5 seconds, 1.5 seconds, etc. Of course, it can also be set to other times, and the embodiments of the present application do not limit this.

[0046] Step 202, according to the space occupied by the point cloud data, divide into several preset specification voxels, and cluster the point cloud data inside each voxel.

[0047] The point cloud data captured by the radar is a set of points in three-dimensional space, and each point cloud data contains information such as the spatial coordinates (x, y, z) of each point. Therefore, the three-dimensional space can be divided into several voxels of a predetermined specification, i.e. the continuous three-dimensional space is discretized into multiple small cubic or cuboid units, thereby efficiently processing the point cloud data.

[0048] The specification of the voxel (i.e. the length of the cube or cuboid) is pre-set, and the selection of this specification can be considered comprehensively according to factors such as the actual application scenario, the density of the point cloud data, and the computing resources.

[0049] The selection of the voxel specification will affect the effect and efficiency of subsequent processing. For example, if faster recognition is required, a larger specification can be set; if more accurate recognition is required, a smaller specification can be set.

[0050] Inside each voxel, the point cloud data is clustered. The purpose of clustering is to gather point cloud data belonging to the same object together, thereby facilitating subsequent feature extraction and pattern recognition.

[0051] Clustering algorithms can group similar points into the same category based on spatial characteristics such as distance and density in point cloud data. When performing clustering within a voxel, the efficiency and effectiveness of the clustering algorithm are improved because the data range is already limited to a smaller space. Clustering algorithms can include K-means clustering, DBSCAN (Density-Based Spatial Clustering Algorithm with Noise), and others.

[0052] Step 203: Process the voxels using feature extraction algorithms and / or pattern recognition algorithms to identify target objects in the target scene.

[0053] The clustered voxels contain point cloud data belonging to the same object, and feature information such as shape features, texture features, and color features (if combined with other sensor data) can be extracted from these point cloud data through feature extraction algorithms and / or pattern recognition algorithms.

[0054] By processing point clouds using feature extraction algorithms to extract information containing the features of target objects, and then using pattern recognition algorithms to analyze these features, the target objects in the target scene can be identified. Pattern recognition algorithms mainly include classifiers and clusterers, which compare the extracted features with a known database of target object features to identify the target objects. Feature extraction algorithms can be divided into various types, including but not limited to: color feature extraction algorithms, texture feature extraction algorithms, and shape feature extraction algorithms.

[0055] Based on the analyzed point cloud data, machine learning or image processing techniques can be used to identify target objects (such as lawn boundaries, obstacles, etc.) and adjust the lawnmower's movement path according to the identification results to achieve precise positioning and operation.

[0056] The target object detection method provided in this application captures point cloud data with LiDAR, performs clustering processing on the point cloud data, and then identifies feature points based on the clustered point cloud. This solves the problem of inaccurate feature point identification or feature point loss during point cloud data processing in the prior art, thereby effectively improving the positioning accuracy and positioning stability of the self-driven device.

[0057] Please refer to the attached document. Figure 3A and attached Figure 3B , attached Figure 3A Used to indicate the original captured point cloud data, attached Figure 3BUsed to indicate the clustered point cloud data, so that the clustering process of the point cloud data can be clearly observed: combining voxelization, clustering, classification and data sampling techniques, the clustered point cloud is made sparse and uniform, but the features of the target object are effectively preserved. It aims to extract meaningful and efficient information from the original LiDAR data, so that users can identify target objects through point cloud data.

[0058] Taking a self-driving lawnmower as an example, this lawnmower is equipped with a high-performance LiDAR, which automatically enters a preset lawn area to scan after startup. The LiDAR collects 3D data of the surrounding environment at a rate of tens of thousands of points per second. After receiving this data, the data processing unit first performs point cloud extraction and filtering to remove noise points caused by uneven ground or weed movement. Subsequently, according to the set 3D voxel space specifications (e.g., 0.01m × 0.01m × 0.01m), the point cloud is divided into multiple voxels, and each voxel is further subdivided into sub-voxels. Using Euclidean clustering, points with close distances within sub-voxels are grouped together to form new voxels. The clustered point cloud is then denoised and homogenized to ensure point cloud quality and reduce data volume, ultimately obtaining point cloud data with a sufficient number of points and continuous features. The voxel data can also be processed using mean and distortion correction algorithms to obtain clean and accurate point cloud data. Finally, based on the processed point cloud data, a machine learning model is used to identify target objects such as road boundaries and obstacles, and the robot's movement path is adjusted according to the distance relationship between the point clouds.

[0059] In an optional embodiment, before performing clustering processing on the point cloud data within each voxel in step 202 of the method provided in this application, steps 301 to 303 are included:

[0060] Step 301: Based on the space occupied by the point cloud data, divide each voxel into several sub-voxels.

[0061] Voxelization involves dividing a 3D space into a series of small cubes (voxels), each containing point cloud data within a certain range. This organizes the originally scattered point cloud into an ordered structure, facilitating subsequent processing. During voxelization, preliminary clustering may be performed, grouping spatially similar points into the same voxel.

[0062] Based on a preset 3D voxel size (dynamically adjusted according to radar performance and outdoor scene), point cloud data is projected onto a voxel grid, with each voxel containing a certain number of points. Points within each voxel are then clustered according to distance to reduce data volume and improve processing efficiency.

[0063] Step 302: By performing Euclidean clustering on the point cloud points contained in each sub-voxel, the sub-voxels occupied by point cloud points within a preset distance in the space are grouped into one class as a single voxel.

[0064] Clustering is the process of grouping points in point cloud data according to a certain similarity criterion. The method provided in this application enables the clustered point cloud data to more clearly reflect the structure of objects in the environment, such as bushes, trees, and flower beds, thereby reducing data redundancy and making subsequent feature extraction and pattern recognition more efficient.

[0065] For example, unorganized point clouds are voxelized to obtain voxel clouds with a fixed resolution. An adjacency graph is established within the voxel space, considering the 26 neighborhoods between voxels. Using this adjacency graph, adjacent voxels are traversed layer by layer from the seed voxel outwards. The similarity distance between adjacent voxels and the current supervoxel is calculated. A flow-constrained clustering algorithm is then used to cluster the voxels into uniform, smooth supervoxels that are attached to the target boundary.

[0066] Secondly, the residual values ​​of all supervoxels are calculated by plane fitting. The curvature is approximated using the residual values ​​to determine the initial seed supervoxels. The supervoxel with the smallest residual value is selected as the initial seed, and a fixed threshold is set to filter other seed supervoxels. Then, feature vectors measuring surface geometry and smoothness are extracted for each supervoxel, and the similarity between adjacent supervoxels is measured by normalized distance values.

[0067] In an optional embodiment, before performing clustering processing on the point cloud data within each voxel in step 202 of the method provided in this application, the method further includes:

[0068] Step 401: Filter the cloud data to obtain the filtered point cloud data.

[0069] The filtering distance threshold is determined based on the total number of point clouds to ensure a certain degree of sparsity and uniformity of the point clouds and obtain a relatively stable number of point clouds.

[0070] Step 402: Perform clustering processing on the filtered point cloud data to obtain at least one voxel.

[0071] This application's embodiments optimize the clustering results by acquiring data from several point clouds, using techniques such as data interpolation and smoothing filtering, making the clustering results more uniform and smooth, thereby improving the accuracy and stability of the positioning.

[0072] In an optional embodiment, after performing clustering processing on the point cloud data within each voxel in step 202 of the method provided in this application, the method further includes:

[0073] Noise in the point cloud data within each voxel is detected, and denoising processing is performed on the point cloud data.

[0074] Noise in point cloud data can be caused by isolated points due to sensor errors, environmental interference, or accidental factors.

[0075] The method provided in this application improves the cleanliness of the data and the accuracy of subsequent processing by detecting noise in point cloud data, removing noise and distortion caused by environmental interference, and retaining point clouds with representativeness and environmental characteristics.

[0076] In an optional embodiment, after performing clustering processing on the point cloud data within each voxel in step 202 of the method provided in this application, the method further includes:

[0077] Step 501: According to the preset classification rules, determine the type of point cloud data within each voxel and determine the priority of point cloud data in each type.

[0078] Due to environmental complexity or the resolution limitations of LiDAR, a single voxel may contain point cloud data from different objects, resulting in multiple types of point cloud data within a voxel: for example, a voxel might simultaneously contain points from the ground, points from leaves, and edge points from distant buildings. These points are grouped into the same voxel because of their spatial proximity, but they actually belong to different objects or parts of the scene.

[0079] Uniform value selection refers to minimizing data redundancy and improving data representativeness while maintaining the overall characteristics of point cloud data.

[0080] Based on segmentation or classification algorithms, point clouds of different categories within a voxel are identified. Priorities are assigned to different categories based on their category and the number of points (i.e., the number of points) within each category. Categories with more points may indicate that the object or scene part occupies a more important position in the environment, and therefore should be given higher priority in subsequent processing.

[0081] This is achieved by selecting representative points within each category according to certain rules (such as uniform sampling, importance sampling, etc.). This method reduces the amount of data processed while retaining sufficient information for subsequent tasks such as environmental understanding and object recognition.

[0082] Step 502: According to the type and priority of the point cloud data, perform mean normalization on the point cloud data within each voxel.

[0083] The method provided in this application embodiment helps to smooth point cloud data by averaging the point cloud data within voxels, reducing fluctuations caused by uneven point cloud density or sensor noise, and also provides more stable and consistent input data for subsequent target object recognition operations.

[0084] In an optional embodiment, after performing clustering processing on the point cloud data within each voxel in step 202 of the method provided in this application, the method includes:

[0085] Step 601: Obtain pose data matching the point cloud data within a preset time period;

[0086] Step 601: Based on the pose data matched with the point cloud data, perform distortion correction on the point cloud data to obtain the feature point cloud after motion distortion compensation.

[0087] By using sensors to collect multiple frames of data for pose integration, the machine's pose at different time periods is obtained. Based on the acquisition time of each point cloud data, the machine's pose is interpolated to calculate the pose of each point cloud. Then, the pose of each point cloud is projected onto the pose of the last point cloud, and finally, the point cloud after distortion correction is obtained. This point cloud data can completely and correctly identify the target object.

[0088] The method provided in this application embodiment is based on the result of point cloud data clustering. First, it processes the noise in the point cloud data and removes the classes that do not meet the point number threshold. Then, it analyzes the categories in each voxel. If there are multiple classes in a voxel, it means that the region has rich features, but it is not stable enough for subsequent extraction of surface features and localization. So it is removed and finally the large class with the point number meets the threshold and the features are continuous is obtained.

[0089] At this point, the number of point clouds is still large. In order to obtain a stable number of point clouds, facilitate real-time operation of the algorithm, and reduce the memory overhead of the algorithm, the point clouds are uniformly sampled according to the priority of category and sorting by the number of points in the category. The more points in a category, the higher the priority. In each category, points are sampled in a loop according to the sorting from the central region to the outer ring until a fixed number of point clouds is obtained.

[0090] After processing, a certain number of representative point clouds with characteristics of the surrounding environment can be obtained, providing strong data support for subsequent positioning.

[0091] In an optional embodiment, step 203 of the method provided in this application, which identifies the target object in the target scene, includes:

[0092] Step 701: Based on several voxels after clustering, extract feature information of the target object. The feature information is used to indicate at least the physical parameters, surface texture, or spatial location of the object.

[0093] Step 702: Identify the target object based on the feature information of the target object.

[0094] Based on clustered point cloud data, feature extraction and pattern recognition techniques are used to identify target objects (such as bushes, walls, etc.). The location information of these target objects can be directly used for lawnmower positioning.

[0095] Pattern recognition is a technology that uses computers to analyze input data and determine its category. In the field of smart lawnmowers, pattern recognition technology is used to match extracted features with predefined models of target objects, thereby identifying objects in the environment. For example, pattern recognition methods may include:

[0096] Example 1: Match the extracted features with the pre-stored templates and find the most similar template as the recognition result.

[0097] Example 2: Using a large amount of labeled data to train a machine learning model, enabling the model to automatically learn and recognize target objects. The machine learning algorithms include support vector machines, decision trees, random forests, and deep learning models (such as convolutional neural networks CNN). Especially when processing image or point cloud data, deep learning models can automatically extract high-level features and perform effective classification and recognition.

[0098] Feature extraction is the process of extracting information useful for object identification from clustered point cloud data. These features may include the object's shape, size, surface texture, spatial location, etc. The feature extraction methods provided in this application include:

[0099] Example 1: By analyzing the coordinate information of points in the point cloud, calculate the geometric shape of the object and extract the geometric features of the voxels (such as boundaries, volume, surface area, etc.).

[0100] Example 2: Use statistical methods, such as histograms and covariance matrices, to describe the distribution characteristics of point clouds and extract the statistical features of voxels.

[0101] Example 3: If the point cloud data contains color or intensity information, the texture features of the objects can also be extracted, which helps to distinguish different types of objects.

[0102] The method provided in this application uses clustered point cloud data as input, extracts key features from the point cloud data using a feature extraction algorithm, matches the extracted features with a predefined model or a trained machine learning model, identifies the target object, and ultimately realizes autonomous navigation, obstacle avoidance, and precise positioning of the self-moving device.

[0103] The method embodiments provided in this application can be executed in a mobile robot or similar computing device. The mobile robot may include one or more processors or programmable logic devices (PLDs) and a memory for storing data. In an exemplary embodiment, the mobile robot may also include a transmission device for communication functions and an input / output device. Those skilled in the art will understand that the mobile robot may also include more or fewer components than those shown above, or have different configurations with equivalent or more functions than those shown above. The memory can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the sensor positioning anomaly determination method in this embodiment of the invention. The processor executes various functional applications and data processing by running the computer program stored in the memory, thereby implementing the above-described method. The memory may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memories.

[0104] In some instances, the memory may further include memory remotely configured relative to the processor, which can be connected to the mobile robot via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device is used to receive or send data via a network. Specific examples of the aforementioned networks may include wireless networks provided by the mobile robot's communication provider.

[0105] In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0106] Please see Figure 4 One embodiment of this application provides a target object detection device 40, comprising:

[0107] The acquisition module 401 collects N frames of point cloud data within a preset time period;

[0108] The processing module 402 divides the point cloud data into several voxels of a preset size according to the space occupied by the point cloud data, and performs cluster processing on the point cloud data within each voxel.

[0109] The identification module 403 processes the voxels using a feature extraction algorithm and / or a pattern recognition algorithm to identify target objects in the target scene.

[0110] In an optional embodiment, the processing module 402 in the target object detection device 40 provided in this application is further configured to:

[0111] Based on the space occupied by the point cloud data, each voxel is divided into several sub-voxels;

[0112] By performing Euclidean clustering on the point cloud points contained in each sub-voxel, sub-voxels occupied by point cloud points within a preset distance in the space are grouped into one class and treated as a single voxel.

[0113] In an optional embodiment, the processing module 402 in the target object detection device 40 provided in this application is further configured to:

[0114] The cloud data is filtered to obtain the filtered point cloud data;

[0115] The filtered point cloud data is clustered to obtain at least one voxel.

[0116] Techniques such as data interpolation and smoothing filtering are used to optimize the clustering results, making them more uniform and smooth, thereby improving the accuracy and stability of positioning.

[0117] In an optional embodiment, the processing module 402 in the target object detection device 40 is further configured to:

[0118] Noise in the point cloud data within each voxel is detected, and denoising processing is performed on the point cloud data.

[0119] Noise is usually isolated points caused by sensor error, environmental interference, or accidental factors.

[0120] In an optional embodiment, the processing module 402 in the target object detection device 40 provided in this application is further configured to:

[0121] Based on several voxels after clustering, feature information of the target object is extracted. The feature information is used to indicate at least the object's physical parameters, surface texture, or spatial location.

[0122] The target object is identified based on its characteristic information.

[0123] In an optional embodiment, the processing module 402 in the target object detection device 40 provided in this application is further configured to:

[0124] Based on the preset classification rules, determine the type of point cloud data within each voxel, and determine the priority of point cloud data in each type;

[0125] The point cloud data within each voxel is averaged based on the type and priority of the point cloud data.

[0126] In an optional embodiment, the processing module 402 in the target object detection device 40 provided in this application is further configured to:

[0127] Obtain pose data matching point cloud data within a preset time period;

[0128] Based on the pose data matched with the point cloud data, distortion correction of the point cloud data is performed to obtain the feature point cloud after motion distortion compensation.

[0129] The target object detection device provided in this application captures point cloud data with lidar, clusters the point cloud data, and then identifies feature points based on the clustered point cloud. This solves the problem of inaccurate feature point identification or feature point loss during point cloud data processing in the prior art, thereby effectively improving the positioning accuracy and positioning stability of the self-driven device.

[0130] For specific limitations regarding the target object detection device of the aforementioned self-driven device, please refer to the limitations of the target object detection method of the self-driven device mentioned above, which will not be repeated here. Each module in the aforementioned target object detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0131] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can perform any step of the target object detection method of the self-driving device described above.

[0132] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A target object detection method characterized by, The method comprises: Collecting N frames of point cloud data within a preset time period; According to the space occupied by the point cloud data, dividing into a plurality of preset specification voxels, and clustering processing the point cloud data inside each voxel; Processing the voxels through a feature extraction algorithm and / or a pattern recognition algorithm to identify target objects in a target scene.

2. The target object detection method according to claim 1, characterized by, The clustering processing the point cloud data inside each voxel comprises: According to the space occupied by the point cloud data, dividing into a plurality of sub-voxels inside each voxel; By performing Euclidean clustering on the point cloud points contained in each sub-voxel, the sub-voxels occupied by the point cloud points within a preset range in the space are classified into a class as a voxel.

3. The device target object detection method of claim 1, wherein, Before the clustering processing the point cloud data inside each voxel, comprising: Filtering the cloud data to obtain filtered point cloud data; Clustering processing the filtered point cloud data to obtain at least one voxel.

4. The device target object detection method according to claim 3, characterized in that, After the clustering processing the point cloud data inside each voxel, further comprising: Detecting the noise points of the point cloud data inside each voxel and performing denoising processing on the point cloud data.

5. The device target object detection method of claim 1, wherein, The method of identifying target objects in a target scene comprises: According to the plurality of voxels after clustering processing, extracting feature information of the target objects, the feature information at least indicating physical parameters, surface texture or spatial position of the objects; According to the feature information of the target objects, identifying the target objects.

6. The device target object detection method of claim 1, wherein, After the clustering processing the point cloud data inside each voxel, further comprising: According to a preset classification rule, determining the type of the point cloud data inside each voxel and the priority of the point cloud data in each type; According to the type of the point cloud data and the priority of the point cloud data, mean value processing the point cloud data inside each voxel.

7. The device target object detection method of claim 1, wherein, After the clustering processing the point cloud data inside each voxel, comprising: Obtaining pose data matched by the point cloud data within a preset time period; According to the pose data matched by the point cloud data, performing distortion correction of the point cloud data to obtain feature point cloud after motion distortion compensation.

8. An object detection device, characterized by, The device comprises: An acquisition module for collecting N frames of point cloud data within a preset time period; A processing module for dividing into a plurality of preset specification voxels according to the space occupied by the point cloud data, and clustering processing the point cloud data inside each voxel; An identification module for processing the voxels through a feature extraction algorithm and / or a pattern recognition algorithm to identify target objects in a target scene.

9. A computer readable storage medium, characterized in that, The storage medium has a computer program stored therein, wherein the computer program is configured to execute the method described in any one of claims 1 to 7 when running.

10. A self-driving device comprising a memory and a processor, characterized in that, The memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method described in any one of claims 1 to 7.

Images