Target detection method and system for improving Point Pill in bulk material scene by using Euclidean clustering

Abstract

The invention relates to a target detection method and system for improving PointPill in a bulk material scene by using Euclidean clustering, and belongs to the technical field of target detection. Through combination of the European clustering algorithm and the PointPill algorithm, the detection precision is effectively improved, and the omission ratio is reduced. According to the method, firstly, a European-style clustering algorithm is used for preprocessing point cloud data, redundant detection frames are recognized and eliminated, and then efficient recognition and positioning of a target are achieved in combination with a detection result of a Point Pill algorithm. The method has good robustness, is suitable for various environmental conditions, improves the detection efficiency, and optimizes the system resource utilization. The method is not only suitable for bulk material pile detection, but also can be widely applied to the fields of automatic driving, robot navigation and the like, and has wide market application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and relates to a target detection method and system in a bulk material scene using Euclidean clustering to improve PointPillar. Background Art

[0002] In the field of bulk material processing, target detection technology is one of the key technologies to achieve bulk material scene operations. The traditional target detection algorithm PointPillar has the following limitations in complex bulk material scenes:

[0003] Redundant detection boxes: When the PointPillar algorithm extracts features and identifies targets from point cloud data, it is easy to generate some redundant detection boxes. These redundant detection boxes may overlap with the actual target, resulting in false detection and thus reducing the detection accuracy.

[0004] Risk of missed detection: The PointPillar algorithm mainly relies on the distribution characteristics of point cloud data for target recognition. In complex bulk material scenes, point cloud data may have problems such as occlusion, missing, and noise, which makes it difficult for the PointPillar model to accurately identify targets with unique shapes (such as bulk material piles, irregular objects, etc.), resulting in the risk of missed detection.

[0005] Insufficient robustness: The PointPillar algorithm has poor adaptability to environmental changes. In situations such as lighting changes, weather changes, and scene occlusions, the detection performance of the PointPillar model may be affected, resulting in reduced detection accuracy.

[0006] In order to overcome the above limitations of the PointPillar algorithm and improve the accuracy and robustness of target detection in complex bulk material scenes, the present invention proposes a target detection method and system for improving PointPillar using Euclidean clustering. Summary of the invention

[0007] In view of this, the object of the present invention is to provide a method and system for target detection in bulk material scenes by using Euclidean clustering to improve PointPillar.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A method for object detection in bulk material scene using Euclidean clustering to improve PointPillar, the method comprising the following steps:

[0010] S1: Use the Euclidean clustering algorithm to cluster the point cloud data and obtain a set of clustering targets;

[0011] S2: Calculate the bounding box (BBOX) of each clustered target, and use the length, width and height information of the known target as a reference;

[0012] S3: Compare the size difference between the BBOX of the clustered target and the known target BBOX, and determine the category of the clustered target according to a preset tolerance threshold;

[0013] S4: Use the PointPillar algorithm to detect the point cloud data and obtain the PointPillar detection result;

[0014] S5: Calculate the overlap between the Euclidean clustering result and the PointPillar detection result;

[0015] S6: Eliminate the PointPillar detection results that overlap with the Euclidean clustering results by 80%;

[0016] S7: For targets that cannot be detected by Euclidean clustering, the PointPillar detection results are retained;

[0017] S8: Output the final target detection result.

[0018] Furthermore, in S3, when the size difference between the BBOX of the clustered target and the known target BBOX is smaller than a preset tolerance threshold, the clustered target is classified as a known target category.

[0019] Furthermore, in S5, the overlapping degree calculation method includes calculating the percentage of the overlapping area of ​​the two bounding boxes to the area of ​​the smaller bounding box.

[0020] Furthermore, in S6, the elimination strategy includes treating PointPillar detection results with an overlap degree greater than a preset threshold as redundant detection frames and removing them.

[0021] A target detection system using Euclidean clustering to improve PointPillar in bulk material scenes, the system comprising:

[0022] The Euclidean clustering recognition module is used to identify the target type based on the length, width and height of the known target;

[0023] PointPillar recognition and overlap calculation module, used to use PointPillar to identify targets and calculate the overlap with Euclidean clustering;

[0024] The target fusion output module is used to remove the targets with an overlap of 80% in the PointPillar results, and the non-overlapping parts are retained;

[0025] The Euclidean clustering recognition module and the PointPillar recognition and overlap calculation module process the input point cloud data respectively and pass the results to the target fusion output module; the target fusion output module receives the results from the Euclidean clustering recognition module and the PointPillar recognition and overlap calculation module, performs fusion processing, and finally outputs the target detection results.

[0026] Furthermore, the Euclidean clustering recognition module includes:

[0027] A clustering unit, used to cluster the input point cloud data using a Euclidean clustering algorithm;

[0028] A size calculation unit, used to calculate the bounding box BBOX of each cluster target;

[0029] The classification unit is used to use the length, width and height information of the known target as a reference and determine the category of the clustered target according to a preset tolerance threshold;

[0030] The clustering unit first clusters the point cloud data and passes the clustering results to the size calculation unit;

[0031] The size calculation unit calculates the bounding box size information of each clustering result and passes this information to the classification unit;

[0032] The classification unit determines the category of the clustered target based on the bounding box size information of the clustered target and the length, width, and height information of the known target, and outputs the category label.

[0033] Furthermore, the PointPillar recognition and overlap calculation module includes:

[0034] PointPillar detection unit, used to detect point cloud data using PointPillar algorithm;

[0035] The overlap calculation unit is used to calculate the overlap between the Euclidean clustering result and the PointPillar detection result;

[0036] The PointPillar detection unit first uses the PointPillar algorithm to detect the point cloud data and passes the detected target bounding box information to the overlap calculation unit;

[0037] The overlap calculation unit receives the PointPillar detection result from the PointPillar detection unit and the clustering result from the Euclidean clustering recognition module, and calculates the overlap degree between the two.

[0038] Furthermore, the target fusion output module includes:

[0039] The redundant detection frame elimination unit is used to eliminate the part of the PointPillar detection result that overlaps with the Euclidean clustering result by 80% according to a preset overlap threshold;

[0040] A result output unit, used to output the final target detection result;

[0041] The redundant detection frame elimination unit receives the PointPillar detection result and overlap information from the PointPillar recognition and overlap calculation module, and eliminates the redundant detection frames;

[0042] The result output unit receives the PointPillar detection result after removing the redundant detection frame from the redundant detection frame elimination unit and the clustering result from the Euclidean clustering recognition module, performs fusion processing, and finally outputs the final target detection result.

[0043] The beneficial effects of the present invention are:

[0044] (1) Preprocessing point cloud data using the Euclidean clustering algorithm can effectively reduce the redundant detection boxes generated by the PointPillar algorithm, thereby improving the accuracy of target detection.

[0045] (2) The method of the present invention can better handle targets with unique or irregular shapes. These targets can be identified through a clustering algorithm and combined with the detection results of PointPillar to reduce the occurrence of missed detections.

[0046] (3) The method of the present invention has strong adaptability to complex bulk material scenarios, can maintain high detection performance under different environmental conditions, and enhances the robustness of the overall system.

[0047] (4) By integrating the results of Euclidean clustering and PointPillar algorithm, the method of the present invention can more efficiently identify and locate the target, thereby improving the detection efficiency.

[0048] (5) By eliminating redundant detection frames, the computational burden of subsequent processing steps is reduced and the utilization of system resources is optimized.

[0049] (6) The method of the present invention is easy to implement in existing point cloud processing and target detection systems and has good compatibility and scalability.

[0050] (7) The method of the present invention is not only applicable to bulk material pile detection, but can also be extended to other application scenarios that require high-precision target detection, such as autonomous driving, robot navigation, industrial automation, etc.

[0051] In summary, the present invention provides a target detection method and system with higher accuracy, lower missed detection rate and stronger robustness in complex bulk material scenarios, which is of great significance for improving the automation level and efficiency of bulk material processing.

[0052] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0054] Figure 1 This is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0055] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0056] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0057] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0058] like Figure 1 As shown, the scheme of the present invention is as follows:

[0059] 1. Input point cloud data.

[0060] This method takes 3D point cloud data as input, obtained from LiDAR or other 3D sensors.

[0061] 2. Use the Euclidean clustering algorithm to classify the point cloud data.

[0062] a) Clustering: Use the Euclidean clustering algorithm to cluster the input point cloud data and obtain a set of clustering target nodes;

[0063] b) Calculate the BBOX of the cluster target: For each cluster, calculate its BBOX. BBOX is the smallest cuboid containing all the points of the cluster, expressed as:

[0064] L, W, H

[0065] c) Get the known target BBOX: Get the length, width and height data of the known target. For example, the length, width and height of the BBOX of a known target are:

[0066] L t , W t , H t

[0067] d) Classification: Assign a category label to each cluster and set a tolerance threshold ∈ to determine whether the BBOX size matches. For example, setting ∈ = 0.1 means that a 10% error is allowed. Compare BBOX size: For each cluster's BBOX, calculate the difference between its length, width, and height and the length, width, and height of the known target BBOX. If the difference is within the threshold range, the cluster is considered to belong to the category. For example, for cluster C, if the following conditions are met, the cluster is classified as a known target category.

[0068] |LL t |≤∈L t ,|WW t |≤∈W t ,|HH t |≤∈H t ,

[0069] Technical innovation: Using the existing target geometry information, such as the typical size of a complete vehicle or bulk material pile, the clustering results are screened to obtain a preliminary target detection frame. This step makes full use of prior knowledge and improves the accuracy of detection.

[0070] 3. Use the PointPillar algorithm to obtain the target detection result, and merge the detection box obtained by Euclidean clustering with the detection result of the PointPillar algorithm. This step aims to combine two different target detection algorithms to give full play to their respective advantages.

[0071] a) Calculate the overlap between the two detection boxes: If the overlap between the two is greater than the specified threshold, they are considered to be the same target.

[0072] Technical innovation: Redundant detection frames are identified by calculating the degree of overlap between the detection frames obtained by Euclidean clustering and the detection results of the PointPillar algorithm.

[0073] b) Remove the parts of the PointPillar detection results that overlap greatly with the Euclidean clustering results to avoid redundant detection.

[0074] Technical innovation: Based on the calculation results of the degree of overlap, the parts of the PointPillar algorithm detection results that highly overlap with the Euclidean clustering results are removed to avoid generating redundant detection frames. This step further improves the accuracy of the detection results.

[0075] 4. For targets that cannot be detected by Euclidean clustering, the detection results of the PointPillar algorithm are retained.

[0076] Through this step, the advantages of the two algorithms can be fully utilized to improve the overall detection coverage.

[0077] Output the final target detection result.

[0078] 5. After the above steps are processed, the final target detection result is output.

[0079] Example 1

[0080] In a bulk material warehouse, the target detection system of the present invention is used to detect the piled materials. The point cloud data input into the system covers an area of ​​10m×10m. The Euclidean clustering recognition module divides the point cloud data into 5 cluster targets, and the length, width and height of each cluster target are: target 1 (3m×2m×1.5m), target 2 (4m×2.5m×2m), target 3 (3.5m×3m×1.8m), target 4 (2m×2m×1m), target 5 (5m×3m×2.5m). The PointPillar recognition and overlap calculation module detected 6 potential targets. By comparing with the Euclidean clustering results, 2 redundant detection frames with an overlap degree of more than 70% were eliminated. Finally, the target fusion output module confirmed 5 material piles, which is exactly the same as the actual number of material piles.

[0081] Example 2

[0082] In mining operations, the system of the present invention is used to detect a 50m×50m ore pile area. Through the Euclidean clustering recognition module, the system identified three main ore piles, whose volumes are: ore pile 1 (8,000 cubic meters), ore pile 2 (12,000 cubic meters), and ore pile 3 (6,000 cubic meters). The PointPillar algorithm detected four potential ore piles. After processing by the target fusion output module, three main ore piles were confirmed, and the boundary of one of the ore piles was corrected. The total volume of ore finally detected was 26,000 cubic meters, which was less than 1% different from the actual measurement value.

[0083] Example 3

[0084] In the cargo stacking area of ​​the port, the system of the present invention is used to detect a 30m×30m cargo stacking area. The Euclidean clustering recognition module identified 10 cargo piles from the point cloud data, and the average size of each cargo pile was 5m×4m×3m. The PointPillar algorithm detected 12 potential cargo piles. The target fusion output module eliminated 2 redundant detection frames with an overlap of more than 60% through comparison, and confirmed the location and size of the 10 cargo piles, which were consistent with the records of the actual cargo piles, with an accuracy rate of 98%.

[0085] The technical innovations of the present invention are mainly reflected in:

[0086] Euclidean clustering is used to make full use of the geometric feature information of known targets.

[0087] The detection results of Euclidean clustering and PointPillar algorithm are integrated to complement each other.

[0088] Effectively eliminate redundant detection boxes that may be generated by the PointPillar algorithm.

[0089] These innovations effectively improve the accuracy and robustness of target detection in complex bulk material scenarios and meet the needs of bulk material processing.

[0090] The PointPillar algorithm often generates some redundant detection boxes, which are prone to false detection and reduce the detection accuracy;

[0091] For some targets with unique shapes (such as bulk material piles), the point cloud data is irregular and varies greatly during the operation process, which makes it difficult for the PointPillar trained model to adapt to the entire operation process. There may be a risk of missed detection, resulting in a reduction in the detection accuracy of the PointPillar algorithm.

[0092] The present invention adopts the Euclidean clustering algorithm to classify the point cloud data, and screens the clustering results according to the geometric features of the known targets (such as length, width and height), which can effectively identify some targets with special shapes and make up for the shortcomings of the PointPillar algorithm.

[0093] The target detection frame obtained by Euclidean clustering is fused with the detection result of PointPillar algorithm, and the overlap degree of the two detection frames is calculated. The part of PointPillar detection result that overlaps with Euclidean clustering result is removed, which can effectively avoid redundant detection and make full use of the geometric feature information of known targets.

[0094] For targets that cannot be detected by Euclidean clustering, the detection results of the PointPillar algorithm are retained to give full play to the respective advantages of the two algorithms, improve the overall detection coverage, and reduce the risk of missed detection.

[0095] Significantly improve the accuracy of target detection: Make full use of the geometric feature information of known targets to accurately identify targets with special shapes. At the same time, through the fusion and screening of detection frames, redundant detection is greatly reduced.

[0096] Improve the robustness of target detection: Combining the PointPillar algorithm based on point cloud distribution characteristics and the Euclidean clustering method based on geometric features can better adapt to various complex bulk material scenes and has strong robustness to interference such as noise and occlusion.

[0097] In summary, the technical solution proposed in the present invention effectively overcomes the limitations of existing target detection algorithms in complex bulk material scenarios, significantly improves the accuracy and robustness of detection, and provides reliable technical support for bulk material processing.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for object detection in bulk material scenes using Euclidean clustering to improve PointPillar, characterized in that: The method comprises the following steps: S1: Use the Euclidean clustering algorithm to cluster the point cloud data and obtain a set of clustering targets; S2: Calculate the bounding box (BBOX) of each clustered target, and use the length, width and height information of the known target as a reference; S3: Compare the size difference between the BBOX of the clustered target and the known target BBOX, and determine the category of the clustered target according to a preset tolerance threshold; S4: Use the PointPillar algorithm to detect the point cloud data and obtain the PointPillar detection result; S5: Calculate the overlap between the Euclidean clustering result and the PointPillar detection result; S6: Eliminate the PointPillar detection results that overlap with the Euclidean clustering results by 80%; S7: For targets that cannot be detected by Euclidean clustering, the PointPillar detection results are retained; S8: Output the final target detection result.

2. According to claim 1, a method for target detection in bulk material scene using Euclidean clustering to improve PointPillar, characterized in that: In S3, when the size difference between the BBOX of the clustered target and the known target BBOX is less than a preset tolerance threshold, the clustered target is classified as a known target category.

3. The target detection method in bulk material scene using Euclidean clustering to improve PointPillar according to claim 1, characterized in that: In S5, the overlapping degree calculation method includes calculating the percentage of the overlapping area of ​​the two bounding boxes to the area of ​​the smaller bounding box.

4. The target detection method in bulk material scene using Euclidean clustering to improve PointPillar according to claim 1, characterized in that: In S6, the elimination strategy includes treating PointPillar detection results with an overlap degree greater than a preset threshold as redundant detection frames and removing them.

5. A target detection system using Euclidean clustering to improve PointPillar in bulk material scenes, characterized by: The system includes: The Euclidean clustering recognition module is used to identify the target type based on the length, width and height of the known target; PointPillar recognition and overlap calculation module, used to use PointPillar to identify targets and calculate the overlap with Euclidean clustering; The target fusion output module is used to remove the targets with an overlap of 80% in the PointPillar results, and the non-overlapping parts are retained; The Euclidean clustering recognition module and the PointPillar recognition and overlap calculation module process the input point cloud data respectively and pass the results to the target fusion output module; the target fusion output module receives the results from the Euclidean clustering recognition module and the PointPillar recognition and overlap calculation module, performs fusion processing, and finally outputs the target detection results.

6. The target detection system using Euclidean clustering to improve PointPillar in bulk material scene according to claim 5, characterized in that: The Euclidean clustering recognition module includes: A clustering unit, used to cluster the input point cloud data using a Euclidean clustering algorithm; A size calculation unit, used to calculate the bounding box BBOX of each cluster target; The classification unit is used to use the length, width and height information of the known target as a reference and determine the category of the clustered target according to a preset tolerance threshold; The clustering unit first clusters the point cloud data and passes the clustering results to the size calculation unit; The size calculation unit calculates the bounding box size information of each clustering result and passes this information to the classification unit; The classification unit determines the category of the clustered target based on the bounding box size information of the clustered target and the length, width, and height information of the known target, and outputs the category label.

7. The target detection system using Euclidean clustering to improve PointPillar in bulk material scenes according to claim 5, characterized in that: The PointPillar recognition and overlap calculation module includes: PointPillar detection unit, used to detect point cloud data using PointPillar algorithm; The overlap calculation unit is used to calculate the overlap between the Euclidean clustering result and the PointPillar detection result; The PointPillar detection unit first uses the PointPillar algorithm to detect the point cloud data and passes the detected target bounding box information to the overlap calculation unit; The overlap calculation unit receives the PointPillar detection result from the PointPillar detection unit and the clustering result from the Euclidean clustering recognition module, and calculates the overlap degree between the two.

8. The target detection system using Euclidean clustering to improve PointPillar in bulk material scenes according to claim 5, characterized in that: The target fusion output module includes: The redundant detection frame elimination unit is used to eliminate the part of the PointPillar detection result that overlaps with the Euclidean clustering result by 80% according to a preset overlap threshold; A result output unit, used to output the final target detection result; The redundant detection frame elimination unit receives the PointPillar detection result and overlap information from the PointPillar recognition and overlap calculation module, and eliminates the redundant detection frames; The result output unit receives the PointPillar detection result after removing the redundant detection frame from the redundant detection frame elimination unit and the clustering result from the Euclidean clustering recognition module, performs fusion processing, and finally outputs the final target detection result.

Images