WO2022205644A1 - 一种目标检测方法、装置、计算机设备和存储介质 - Google Patents

一种目标检测方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2022205644A1
WO2022205644A1 PCT/CN2021/103186 CN2021103186W WO2022205644A1 WO 2022205644 A1 WO2022205644 A1 WO 2022205644A1 CN 2021103186 W CN2021103186 W CN 2021103186W WO 2022205644 A1 WO2022205644 A1 WO 2022205644A1
Authority
WO
WIPO (PCT)
Prior art keywords
point cloud
cloud data
fragmented
radar
object detection
Prior art date
Application number
PCT/CN2021/103186
Other languages
English (en)
French (fr)
Inventor
付万增
王哲
石建萍
Original Assignee
上海商汤临港智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海商汤临港智能科技有限公司 filed Critical 上海商汤临港智能科技有限公司
Publication of WO2022205644A1 publication Critical patent/WO2022205644A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2200/00—Indexing scheme for image data processing or generation, in general
    • G06T2200/32—Indexing scheme for image data processing or generation, in general involving image mosaicing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10028—Range image; Depth image; 3D point clouds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • the present disclosure relates to the field of computer technology, and relates to a target detection method, apparatus, computer device and storage medium.
  • the radar perception algorithm provides stable obstacle perception results for the automatic driving system, so the radar perception algorithm has become one of the indispensable technologies for the automatic driving system.
  • Radar perception algorithms need to process a huge amount of point cloud input data and provide accurate and real-time perception results, which poses a huge challenge to the complexity and hardware performance of radar perception algorithms.
  • radar perception algorithms are often based on deep neural networks with huge parameters, so it is difficult for radar perception algorithms to meet the real-time requirements of autonomous driving systems.
  • Embodiments of the present disclosure provide at least one object detection method, apparatus, computer device, and storage medium.
  • an embodiment of the present disclosure provides a target detection method, including:
  • the first fragmented point cloud data is collected every time the radar rotates by a preset angle point cloud data; the preset angle is less than 360 degrees;
  • the point cloud data to be spliced is the first fragmented point cloud data previously output by the radar, and the Part of the point cloud data adjacent to the first fragmented point cloud data currently output by the radar;
  • part of the point cloud data in the point cloud data of the first fragment obtained previously is spliced with the point cloud data of the first fragment obtained currently, so as to obtain comprehensive information near the splicing place. Therefore, it is possible to Overcome the problem of missed detection at the point cloud data segmentation, and improve the detection accuracy.
  • the point cloud data collected by the radar can be processed into pieces, which can reduce the amount of data processed by the radar in a single time, thereby improving the detection efficiency. When applied to automatic driving systems, it can meet the requirements of real-time detection results.
  • the extraction of point cloud data to be spliced from the first fragmented point cloud data previously output by the radar includes:
  • the first boundary is the point cloud boundary formed when the radar collects the first fragmented point cloud data, and the first boundary is perpendicular to the direction of rotation of the radar;
  • the point cloud data whose distance from the first boundary is less than the first preset distance is screened, and the point cloud data obtained by screening is used as the point cloud data to be spliced.
  • the point cloud data that is connected to the point cloud data of the first fragment that is currently acquired is selected as the point cloud data to be spliced, which can complement the point cloud data.
  • the point cloud data at the point cloud segmentation can obtain comprehensive information near the splicing point, thereby overcoming the problem of missed detection at the point cloud data segmentation.
  • the splicing of the point cloud data to be spliced with the first fragmented point cloud data currently output to obtain the second fragmented point cloud data includes:
  • the second boundary is the boundary point cloud data corresponding to when the radar starts to collect the first fragmented point cloud data, and the second boundary is the boundary is perpendicular to the direction of rotation of said radar;
  • the point cloud data to be spliced is spliced on the second boundary side of the currently output first segmented point cloud data to obtain second segmented point cloud data.
  • the point cloud data to be spliced can be accurately spliced onto the currently output first fragmented point cloud data to form a second fragment that can cover the currently acquired first fragmented point cloud data.
  • the fragmented point cloud data can solve the problem of missed detection at the segmentation of the first fragmented point cloud data currently output.
  • the target detection result includes at least one object detection frame and a confidence level of each object detection frame
  • the method further includes:
  • Object detection frames other than the object detection frame corresponding to the highest confidence level in the first detection frame group are eliminated.
  • object detection frames other than the object detection frame corresponding to the highest confidence level in the first detection frame group are eliminated, which effectively reduces the false detection rate and improves the detection accuracy.
  • the object detection frame corresponding to the point cloud data of the first fragment that was previously output and the object detection frame corresponding to the point cloud data of the second fragment are obtained by obtaining at least part of the frame.
  • the overlapping first detection frame group including:
  • the ratio of the area of the overlapping area of the at least partially overlapping two object detection frames to the area of any one of the at least partially overlapping two object detection frames is greater than the preset value, set the The at least partially overlapping two object detection frames serve as the first detection frame group.
  • the object detection frame with a high probability of belonging to the same object can be used as a first detection frame group through the overlapping area of the two overlapping object detection frames, so that the object detection frame with low confidence in it can be eliminated, and the The object detection frame with the highest confidence is used as the final object detection frame of the corresponding object, avoiding the situation where there are two object detection frames for the same object, which can effectively reduce the false detection rate and improve the detection accuracy; in addition, two objects with a small overlap area
  • the detection frame has a high probability of belonging to different objects, and there is no need to process the object detection frame in this case, which further improves the detection accuracy.
  • the target detection result includes at least one object detection frame and a confidence level of each object detection frame
  • the method further includes:
  • the third boundary is perpendicular to the rotation direction of the radar
  • the confidence of the first target detection frame is modified based on a preset attenuation scale factor, the distance from the center point of the first target detection frame to the third boundary, and the second preset distance.
  • this embodiment can be used to convert the object detection frame closer to the boundary.
  • the confidence level is corrected to a confidence level consistent with the actual situation, so that a more accurate target detection result on the third boundary side can be obtained.
  • an embodiment of the present disclosure further provides a target detection device, including:
  • the acquisition part is configured to acquire the first fragmented point cloud data currently output by the radar and the first fragmented point cloud data previously output by the radar; the first fragmented point cloud data is one fragment per rotation of the radar. point cloud data collected at a preset angle; the preset angle is less than 360 degrees;
  • the extraction part is configured to extract point cloud data to be spliced from the first fragmented point cloud data output by the radar last time; the point cloud data to be spliced is the first fragmented point output by the radar last time In the cloud data, part of the point cloud data adjacent to the first fragmented point cloud data currently output by the radar;
  • the splicing part is configured to splicing the point cloud data to be spliced with the first fragmented point cloud data currently output to obtain the second fragmented point cloud data;
  • the detection part is configured to detect the second fragmented point cloud data to obtain a target detection result.
  • the extracting part is configured to determine the first boundary of the point cloud data of the first fragment that was output last time; the first boundary is for the radar to collect the first fragment.
  • the point cloud boundary formed when the point cloud data ends, and the first boundary is perpendicular to the rotation direction of the radar;
  • the point cloud data whose distance from the first boundary is less than the first preset distance is screened, and the point cloud data obtained by screening is used as the point cloud data to be spliced.
  • the splicing part is configured to determine the second boundary of the point cloud data of the first fragment that is currently output; the second boundary is when the radar starts to collect the first fragment.
  • the boundary point cloud data corresponding to the point cloud data, and the second boundary is perpendicular to the rotation direction of the radar;
  • the point cloud data to be spliced is spliced on the second boundary side of the currently output first segmented point cloud data to obtain second segmented point cloud data.
  • the target detection result includes at least one object detection frame and a confidence level of each object detection frame
  • the detection part is further configured to detect the second fragmented point cloud data and obtain the target detection result, from the object detection frame corresponding to the first fragmented point cloud data output previously, and the first fragmented point cloud data.
  • the object detection frame corresponding to the two-slice point cloud data obtain at least a partially overlapping first detection frame group;
  • Object detection frames other than the object detection frame corresponding to the highest confidence level in the first detection frame group are eliminated.
  • the detection part is configured to be from the object detection frame corresponding to the first fragmented point cloud data output previously, and the object detection frame corresponding to the second fragmented point cloud data.
  • the object detection frame two at least partially overlapping object detection frames are obtained; in the overlapping area area of the at least partially overlapping two object detection frames, any one of the at least partially overlapping two object detection frames is described
  • the at least partially overlapping two object detection frames are used as one of the first detection frame groups.
  • the target detection result includes at least one object detection frame and a confidence level of each object detection frame; the detection part is further configured to analyze the second fragmented point cloud data.
  • Carry out detection after obtaining the target detection result, determine the third boundary of the second fragmented point cloud data; the third boundary is perpendicular to the rotation direction of the radar; from the second fragmented point cloud data corresponding In the object detection frame, the first target detection frame whose distance from the third boundary is less than the second preset distance is screened;
  • the confidence of the first target detection frame is modified based on a preset attenuation scale factor, the distance from the center point of the first target detection frame to the third boundary, and the second preset distance.
  • embodiments of the present disclosure further provide a computer device, including: a processor, a memory, and a bus, where the memory stores machine-readable instructions executable by the processor, and when the computer device runs, the processing A bus communicates between the processor and the memory, and when the machine-readable instructions are executed by the processor, the first aspect or the steps in any possible implementation manner of the first aspect are performed.
  • embodiments of the present disclosure further provide a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and the computer program is executed by a processor to execute the first aspect, or any one of the first aspect. steps in one possible implementation.
  • an embodiment of the present disclosure provides a computer program, including computer-readable code, when the computer-readable code is executed in an electronic device, a processor in the electronic device executes the above-mentioned first aspect, or Steps in any possible implementation of the first aspect.
  • the target detection method, device, computer equipment, and storage medium can obtain the first fragmented point cloud data currently output by the radar and the first fragmented point cloud data previously output by the radar;
  • the slice point cloud data is the point cloud data collected by the radar every time it rotates a preset angle; the preset angle is less than 360 degrees; the point cloud data to be spliced is extracted from the first segment point cloud data output by the radar before;
  • the spliced point cloud data is the part of the point cloud data adjacent to the first segmented point cloud data currently output by the radar in the first segmented point cloud data previously output by the radar;
  • the point cloud data of one fragment is spliced to obtain the point cloud data of the second fragment; the point cloud data of the second fragment is detected to obtain the target detection result, which is different from the existing technology when the real-time output detection result is satisfied.
  • the splicing point Compared with the low accuracy of the detection result of the splicing point, it splices part of the point cloud data in the point cloud data of the first shard obtained previously with the point cloud data of the first shard currently obtained, so as to obtain a comprehensive picture near the splicing point. Therefore, the problem of missed detection at the point cloud data segmentation can be overcome, and the detection accuracy can be further improved.
  • the point cloud data collected by the radar can be processed into pieces, which can reduce the amount of data processed by the radar in a single time, thereby improving the detection efficiency. When applied to automatic driving systems, it can meet the requirements of real-time detection results.
  • FIG. 1 shows a flowchart of a target detection method provided by an embodiment of the present disclosure
  • FIG. 2 shows a schematic diagram of radar fragmentation outputting first fragmented point cloud data provided by an embodiment of the present disclosure
  • FIG. 3 shows a schematic flowchart of obtaining point cloud data of a second fragment provided by an embodiment of the present disclosure
  • FIG. 4 shows a flowchart of excluding duplicate detection results in target detection results after splicing provided by an embodiment of the present disclosure
  • FIG. 5 shows a schematic diagram of the distribution of object detection frames in the first fragmented point cloud data and the second fragmented point cloud data provided by an embodiment of the present disclosure
  • FIG. 6 shows a flowchart of correcting the confidence of the first target to the measuring frame provided by an embodiment of the present disclosure
  • FIG. 7 shows a schematic diagram of a target detection apparatus provided by an embodiment of the present disclosure
  • FIG. 8 shows a schematic structural diagram of a computer device provided by an embodiment of the present disclosure.
  • references herein to "a plurality or several” means two or more.
  • "And/or" which describes the association relationship of the associated objects, means that there can be three kinds of relationships, for example, A and/or B, which can mean that A exists alone, A and B exist at the same time, and B exists alone.
  • the character "/" generally indicates that the associated objects are an "or" relationship.
  • the application scenario of the target detection method disclosed in the embodiment of the present disclosure is firstly introduced, and the working process of the radar in the application scenario is firstly introduced:
  • the point cloud data collected by the radar is to emit laser light through the radar transmitter, encounter the reflection of the obstacle, and then receive the laser light from the radar receiver.
  • the farthest distance that the laser encounters an obstacle and returns to be accepted by the receiver determines the farthest range the radar can detect.
  • Radars are generally equipped with multiple pairs of transmitters and receivers, with different transmitters and receivers emitting or receiving laser light at different angles.
  • the radar may include a mechanical lidar.
  • the transmitter of mechanical lidar can also rotate along a fixed rotation axis while emitting and receiving laser light to obtain 360-degree omnidirectional point cloud data information.
  • the whole frame of data can be sent to subsequent modules (such as neural network modules) through the data port.
  • the final presented point cloud data is a collection of point cloud data in a cylindrical area with the radar rotation center as the center and the farthest detection range as the radius.
  • the height of the cylinder is determined by the maximum angle between different transmitters and receivers. Decide.
  • the range of point cloud data that can be collected in practical applications is smaller than that of a regular cylinder.
  • the embodiments of the present disclosure involve
  • the detection range of the radar is considered to be a cylinder.
  • the target detection method provided by the embodiment of the present invention can be applied to an application scenario of automatic driving.
  • the radar is usually installed on the roof of the vehicle, because the neural network model for processing point cloud data is good at processing information in a rectangular area.
  • the original point cloud data collected by the radar needs to be further cropped to obtain a cuboid with the radar as the center.
  • the high side is perpendicular to the ground.
  • the radar collects point cloud data and outputs information on the types and positions of obstacles in the cuboid area where the point cloud data is located.
  • Use the neural network model to detect the point cloud data in the cuboid area, and output the detection results in the cuboid area.
  • the above cuboid is the following object detection frame.
  • the point cloud data collected by the radar rotates 360 degrees
  • the point cloud data is processed through the neural network model, and the 360 degree detection result is finally output.
  • the device includes, for example, a terminal device or a server or other processing device, and the terminal device can be a user equipment (User Equipment, UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (Personal Digital Assistant, PDA) , handheld devices, computing devices, in-vehicle devices, wearable devices, etc.
  • the object detection method may be implemented by the processor calling computer-readable instructions stored in the memory.
  • the target detection method provided by the embodiment of the present disclosure will be described below by taking the execution subject as a computer device as an example.
  • FIG. 1 is a flowchart of a target detection method provided by an embodiment of the present disclosure.
  • the method includes steps S101 to S104 , wherein:
  • S101 Obtain the first fragmented point cloud data currently output by the radar and the first fragmented point cloud data previously output by the radar.
  • the first fragmented point cloud data is the point cloud data collected by the radar every time the radar rotates by a preset angle.
  • the preset angle is less than 360 degrees.
  • the preset angle can be customized according to the calculation amount of the neural network model and the complexity of the test object in the application scenario. limited.
  • the point cloud data collected by the mechanical lidar has time series information, that is, by rotating the laser transmitter and receiver 360 degrees, the point cloud data can be continuously collected.
  • FIG. 2 which is a schematic diagram of the radar slice outputting point cloud data of the first slice, where every 90 degrees is a sub-cycle, a piece of point cloud data of the first slice is output.
  • the point cloud data to be spliced is part of the point cloud data adjacent to the first fragmented point cloud data currently output by the radar in the first fragmented point cloud data previously output by the radar.
  • the first preset distance can be customized according to experience values in different application scenarios, and is not limited here.
  • the first preset distance can be defined according to the size data of the detection object, for example, the vehicle length is 6 meters, the first preset distance can be selected from data greater than 6 meters, and the data can be based on experience.
  • the value and the actual size data of the detection object are comprehensively considered as a trade-off result, which is not limited here.
  • FIG. 3 is a schematic flowchart of obtaining the point cloud data of the second fragment.
  • the first boundary 311 of the first fragmented point cloud data 31 is determined, and the first boundary 311 of the first fragmented point cloud data 31 is determined.
  • point cloud data whose distance from the first boundary 311 is smaller than the first preset distance are screened, and the screened point cloud data is used as the point cloud data to be spliced 312 .
  • S103 Splicing the point cloud data to be spliced with the point cloud data of the first fragment that is currently output to obtain the point cloud data of the second fragment.
  • the second boundary of the first fragmented point cloud data currently output is determined; the second boundary is the boundary point cloud data corresponding to when the radar starts to collect the first fragmented point cloud data, and the second boundary is the same as that of the first fragmented point cloud data.
  • the rotation direction of the radar is vertical; the point cloud data to be spliced is spliced on the second boundary side of the first segmented point cloud data currently output to obtain the second segmented point cloud data.
  • the above-mentioned use of the second boundary can accurately splicing the point cloud data to be spliced on the currently output first shard point cloud data to form a second shard point that can cover the currently acquired first shard point cloud data
  • Cloud data can solve the problem of missed detection at the segmentation of the first fragmented point cloud data currently output.
  • the second boundary 321 of the first fragmented point cloud data 32 is determined.
  • the point cloud data 312 to be spliced exists and The same boundary of the first fragmented point cloud data 31, that is, the first boundary 311, the first boundary 311 side of the point cloud data 312 to be spliced is spliced to the second boundary 321 side of the first fragmented point cloud data 32 to obtain The second fragmented point cloud data 33 .
  • S104 Detect the second fragmented point cloud data to obtain a target detection result.
  • the second fragmented point cloud data can be detected by using a neural network model to obtain a target detection result output by the neural network model.
  • the first fragmented point cloud data in the first sub-cycle collected by the radar output can also be detected by using the neural network model, and the target detection result output by the neural network model can be obtained.
  • the target detection results of a whole cycle of the radar output by the neural network model can be spliced to obtain the target detection results of all angles.
  • Another implementation is to splicing the previous target detection result output by the neural network model with the current target detection result to obtain the target detection result of a partial angle.
  • FIG. 4 is a flowchart of eliminating duplicate detection results in the spliced target detection results.
  • S401 Obtain an at least partially overlapping first detection frame group from the object detection frame corresponding to the point cloud data of the first fragment that was previously output and the object detection frame corresponding to the point cloud data of the second fragment.
  • the object detection frame corresponding to the point cloud data of the first fragment that was previously output can be determined in the following ways:
  • Method 1 In the case where the first piece of point cloud data outputted in the previous time is the first piece of point cloud data output by the radar, use the neural network model to detect the first piece of point cloud data outputted in the previous time, and obtain the previous piece of point cloud data.
  • the object detection frame corresponding to the output point cloud data of the first fragment.
  • Method 2 In the case where the first piece of point cloud data outputted in the previous time is the first piece of point cloud data output by the radar, use the neural network model to detect the second piece of point cloud data outputted in the previous The point cloud data of the first fragment is a part of the point cloud data of the second fragment output previously), and the obtained object detection frame is used as the object detection frame corresponding to the point cloud data of the first fragment output previously.
  • two object detection frames that at least partially overlap can be obtained from the object detection frame corresponding to the point cloud data of the first fragment that was previously output, and the object detection frame corresponding to the point cloud data of the second fragment;
  • the ratio of the area of the overlapping area of the at least partially overlapping two object detection frames to the area of any one of the at least partially overlapping two object detection frames is greater than the preset value, the at least partially overlapping two The object detection frame is used as a first detection frame group.
  • the two object detection frames with smaller overlapping area are likely to belong to different objects, and it is not necessary to process the object detection frames in this case.
  • the preset value can be customized according to the experience value in different application scenarios, which is not limited here.
  • the distribution positions of the object detection frames in the first fragmented point cloud data 50 are determined, wherein the neural network model identifies that there are object detection frames 511 and object detection frames in the point cloud data 51 to be spliced. 512 and the object detection frame 523 are incomplete detection frames, the neural network model can predict the remaining part of the detection frame, that is, the dashed part in the object detection frame 511, the dashed part in the object detection frame 512, and the object detection frame 523. The dashed part in .
  • Two object detection frames that at least partially overlap can be obtained from the object detection frame in the first fragmented point cloud data 50 previously output, and the object detection frame corresponding to the second fragmented point cloud data 52, wherein the object detection frame 511 and object detection frame 521, object detection frame 512 and object detection frame 522, object detection frame 513 and object detection frame 523, object detection frame 514 and object detection frame 524, and object detection frame 511 and object detection frame 514 are partially overlapping
  • Two object detection frames can determine the object detection frame 511 and the object detection frame 521, the object detection frame 512 and the object detection frame 522, the object detection frame 513 and the object detection frame 523, and the object detection frame 514 and the object detection frame.
  • the object detection frame 524 satisfies the condition that the area of the overlapping area of the two object detection frames and the area of any object detection frame in the at least partially overlapping two object detection frames are greater than the preset value.
  • the above four groups of object detection frames can be as the first detection frame group.
  • the ratio of the area of the overlapping area of the object detection frame 511 and the object detection frame 514 to the area of either the object detection frame 511 or the object detection frame 514 is less than or equal to the preset value, so the object detection frame 511 and the object detection frame 514 cannot be used as the first A detection frame group.
  • S402 Eliminate object detection frames other than the object detection frame corresponding to the highest confidence in the first detection frame group.
  • the target detection result includes at least one object detection frame and a confidence level of each object detection frame.
  • the non-maximum value suppression algorithm based on the first detection frame group obtained in step S401 and the confidence level of each object detection frame in the first detection frame group, determine the one with the highest confidence in the first object detection frame group Object detection frame, the object detection frame with the highest confidence is retained, and the remaining object detection frames in the first detection frame group are eliminated.
  • the object detection frame with low confidence is eliminated, and the object detection frame with the highest confidence is used as the final object detection frame of the corresponding object, so as to avoid the situation that there are two object detection frames for the same object, It can effectively reduce the false detection rate and improve the detection accuracy.
  • the spliced target detection results are processed, and the confidence level of the first target detection frame can also be corrected before the duplicate detection results in the spliced target detection results are eliminated.
  • FIG. 6 is a flow chart of correcting the confidence of the first target on the frame, wherein:
  • S601 Determine a third boundary of the second fragmented point cloud data; the third boundary is perpendicular to the rotation direction of the radar.
  • the third boundary of the second fragmented point cloud data may be the boundary 331 and the boundary 332 in FIG. 3 .
  • S602 Screen the first object detection frame whose distance from the third boundary is smaller than the second preset distance from the object detection frame corresponding to the second fragmented point cloud data.
  • the second preset distance can be customized according to the experience value in different application scenarios, which is not limited here.
  • S603 Correct the confidence of the first target detection frame based on the preset attenuation scale factor, the distance from the center point of the first target detection frame to the third boundary, and the second preset distance.
  • the preset attenuation scale factor may take a value in the range of 0 to 1.
  • new_score old_score ⁇ (1-scale_factor ⁇ max(0,(D-d)/D)) formula (1);
  • new_score represents the corrected confidence of the first target detection frame
  • old_score represents the confidence of the first target detection frame output by the neural network model
  • scale_factor represents the attenuation scale factor
  • D represents the second preset distance
  • d represents the distance from the center of the first target detection frame to the third boundary
  • max(0, (D-d)/D) represents the maximum value between 0 and (D-d)/D.
  • the data obtained by detecting the first fragmented point cloud data can also be analyzed.
  • the point cloud detection results ie the first point cloud detection result and the second point cloud detection result
  • de-duplication processing is directly subjected to de-duplication processing.
  • the first fragmented point cloud data currently output by the radar is detected to obtain the first point cloud detection result; the first fragmented point cloud data previously output by the radar is detected to obtain the second point cloud Detection result; from the object detection frame corresponding to the first point cloud detection result and the object detection frame corresponding to the second point cloud detection result, obtain at least a partially overlapping second detection frame group; divide the second detection frame group with the highest Object detection frames other than the object detection frame corresponding to the confidence level are eliminated.
  • the confidence degree of the object detection frame also needs to be corrected.
  • the fourth boundary of the point cloud data of the first fragment is determined; the fourth boundary is perpendicular to the rotation direction of the radar; The second target detection frame whose distance is less than the third preset distance; the second target detection frame is detected based on the preset attenuation scale factor, the distance from the center point of the second target detection frame to the fourth boundary point, and the third preset distance.
  • the confidence level of the box is corrected.
  • the point cloud data to be spliced obtained from the point cloud data of the first segment outputted previously is spliced to the first segment currently output by using the point cloud data of the first segment output by the radar segment.
  • new fragmented point cloud data is obtained, that is, the second fragmented point cloud data.
  • the obtained second fragmented point cloud data includes all the point clouds where the currently output first fragmented point cloud data is spliced. Therefore, by splicing part of the point cloud data in the point cloud data of the first fragment obtained previously with the point cloud data of the first fragment obtained currently, the global information near the point cloud data segmentation can be obtained.
  • the point cloud data collected by the radar can be processed into pieces, which can reduce the amount of data processed by the radar in a single time, thereby improving the detection efficiency.
  • it can meet the requirements of real-time acquisition of detection results.
  • the writing order of each step does not mean a strict execution order but constitutes any limitation on the implementation process, and the execution order of each step should be based on its function and possible intrinsic Logical OK.
  • the embodiment of the present disclosure also provides a target detection device corresponding to the target detection method.
  • the implementation can refer to the implementation of the method.
  • the device includes: an acquisition part 701 , an extraction part 702 , a splicing part 703 , and a detection part 704 ; wherein,
  • the acquisition part 701 is configured to acquire the first fragmented point cloud data currently output by the radar and the first fragmented point cloud data previously output by the radar; the first fragmented point cloud data is the data for each rotation of the radar. Point cloud data collected at a preset angle; the preset angle is less than 360 degrees;
  • the extraction part 702 is configured to extract the point cloud data to be spliced from the first fragmented point cloud data output by the radar last time; the point cloud data to be spliced is the first fragment outputted by the radar last time In the point cloud data, part of the point cloud data adjacent to the first fragmented point cloud data currently output by the radar;
  • the splicing part 703 is configured to splicing the point cloud data to be spliced with the first fragmented point cloud data currently output to obtain the second fragmented point cloud data;
  • the detection part 704 is configured to detect the second fragmented point cloud data to obtain a target detection result.
  • the extraction part 702 is configured to determine a first boundary of the first segmented point cloud data output last time; the first boundary is the first segment collected by the radar.
  • the point cloud data whose distance from the first boundary is less than the first preset distance is screened, and the point cloud data obtained by screening is used as the point cloud data to be spliced.
  • the splicing part 703 is configured to determine a second boundary of the first fragmented point cloud data currently output; the second boundary is when the radar starts to collect the first point cloud data.
  • the boundary point cloud data corresponding to the point cloud data, and the second boundary is perpendicular to the rotation direction of the radar;
  • the point cloud data to be spliced is spliced on the second boundary side of the currently output first segmented point cloud data to obtain second segmented point cloud data.
  • the target detection result includes at least one object detection frame and a confidence level of each object detection frame
  • the detection part 704 is further configured to detect the second fragmented point cloud data and obtain the target detection result, from the object detection frame corresponding to the first fragmented point cloud data previously output, and the object detection frame. In the object detection frame corresponding to the second fragmented point cloud data, obtain at least a partially overlapping first detection frame group;
  • Object detection frames other than the object detection frame corresponding to the highest confidence level in the first detection frame group are eliminated.
  • the detection part 704 is configured to be from the object detection frame corresponding to the point cloud data of the first fragment that was previously output, and the frame corresponding to the point cloud data of the second fragment.
  • the object detection frame two at least partially overlapping object detection frames are obtained; in the overlapping area area of the at least partially overlapped two object detection frames, any one of the at least partially overlapping two object detection frames is When the ratio of the areas of the object detection frames is greater than a preset value, the at least partially overlapping two object detection frames are used as one of the first detection frame groups.
  • the target detection result includes at least one object detection frame and the confidence level of each object detection frame; the detection part 704 is further configured to The data is detected, and after the target detection result is obtained, the third boundary of the second fragmented point cloud data is determined; the third boundary is perpendicular to the rotation direction of the radar; from the second fragmented point cloud data The corresponding object detection frame is screened for a first target detection frame whose distance from the third boundary is less than a second preset distance;
  • the confidence of the first target detection frame is modified based on a preset attenuation scale factor, the distance from the center point of the first target detection frame to the third boundary, and the second preset distance.
  • FIG. 8 a schematic structural diagram of a computer device provided by an embodiment of the present application includes:
  • Processor 81 memory 82 and bus 83 .
  • the memory 82 stores machine-readable instructions executable by the processor 81, and the processor 81 is configured to execute the machine-readable instructions stored in the memory 82.
  • the processor 81 Perform the following steps:
  • the first fragmented point cloud data is the point collected every time the radar rotates by a preset angle cloud data; the preset angle is less than 360 degrees;
  • the point cloud data to be spliced is spliced with the currently output first fragmented point cloud data to obtain the second fragmented point cloud data;
  • the second fragmented point cloud data is detected to obtain a target detection result.
  • the above-mentioned memory 82 includes a memory 821 and an external memory 822; the memory 821 here is also called an internal memory, and is configured to temporarily store the operation data in the processor 81 and the data exchanged with the external memory 822 such as the hard disk.
  • the processor 81 passes the memory 821 Data is exchanged with the external memory 822.
  • the processor 81 communicates with the memory 82 through the bus 83, so that the processor 81 executes the execution instructions mentioned in the above method embodiments.
  • Embodiments of the present disclosure further provide a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, the steps of the target detection method described in the above method embodiments are executed.
  • the storage medium may be a volatile or non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure further provides a computer program product, where the computer program product carries program codes, and the instructions included in the program codes can be used to execute the steps of the target detection method described in the foregoing method embodiments, and reference may be made to the foregoing method.
  • the above-mentioned computer program product can be realized by means of hardware, software or a combination thereof.
  • the computer program product may be embodied as a computer storage medium, and in another optional embodiment, the computer program product may be embodied as a software product, such as a software development kit (Software Development Kit, SDK), etc. Wait.
  • the units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • each functional part in each embodiment of the present disclosure may be integrated into one processing part, or each part may exist physically alone, or two or more parts may be integrated into one part.
  • the functions, if implemented as functional parts of software and sold or used as separate products, may be stored in a processor-executable non-volatile computer-readable storage medium.
  • the technical solutions of the embodiments of the present disclosure are essentially or contribute to the prior art or parts of the technical solutions may be embodied in the form of software products, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present disclosure.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes .
  • the present disclosure provides a method, device, computer equipment and storage medium for target detection, wherein the method includes: acquiring first fragmented point cloud data currently output by the radar and first fragmented point cloud data previously output by the radar ;
  • the first fragmented point cloud data is the point cloud data collected by the radar every time the radar rotates a preset angle; the preset angle is less than 360 degrees; from the first fragmented point cloud data output by the radar before, extract the point to be spliced Cloud data;
  • the point cloud data to be spliced is the part of the point cloud data adjacent to the first segmented point cloud data currently output by the radar in the first segmented point cloud data previously output by the radar;
  • the point cloud data to be spliced with The currently output point cloud data of the first fragment is spliced to obtain the point cloud data of the second fragment; the point cloud data of the second fragment is detected to obtain the target detection result.
  • the method is achieved by comparing part of the point cloud data in the first fragmented point cloud data acquired previously with the currently acquired first point cloud data.
  • the fragmented point cloud data are spliced to obtain comprehensive information near the splicing point. Therefore, the problem of missed detection at the segmentation point of the point cloud data can be overcome, and the detection accuracy can be further improved.
  • the point cloud data collected by the radar can be processed into pieces, which can reduce the amount of data processed by the radar in a single time, thereby improving the detection efficiency. When applied to automatic driving systems, it can meet the requirements of real-time detection results.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Quality & Reliability (AREA)
  • Health & Medical Sciences (AREA)
  • Geometry (AREA)
  • Multimedia (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

本公开提供了一种目标检测的方法、装置、计算机设备和存储介质，其中，该方法包括：获取雷达当前输出的第一分片点云数据以及雷达前一次输出的第一分片点云数据；第一分片点云数据为雷达每旋转一个预设角度所采集到的点云数据；预设角度小于360度；从雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；待拼接点云数据为雷达前一次输出的第一分片点云数据中、与雷达当前输出的第一分片点云数据相邻的部分点云数据；将待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；对第二分片点云数据进行检测，得到目标检测结果。

Description

一种目标检测方法、装置、计算机设备和存储介质
相关申请的交叉引用
本公开基于申请号为202110332125.4、申请日为2021年03月29日、申请名称为“一种目标检测方法、装置、计算机设备和存储介质”的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本公开作为参考。
技术领域
本公开涉及计算机技术领域，涉及一种目标检测方法、装置、计算机设备和存储介质。
背景技术
目前，雷达感知算法为自动驾驶***提供了稳定的障碍物感知结果，因此雷达感知算法已经成为自动驾驶***不可或缺的技术之一。雷达感知算法需要处理数量庞大的点云输入数据，并提供准确实时的感知结果，这对雷达感知算法的复杂度和硬件性能提出了巨大的挑战。另外，雷达感知算法往往基于参数量巨大的深度神经网络，因此雷达感知算法很难满足自动驾驶***实时性的要求。
发明内容
本公开实施例至少提供一种目标检测方法、装置、计算机设备和存储介质。
第一方面，本公开实施例提供了一种目标检测方法，包括：
获取雷达当前输出的第一分片点云数据以及所述雷达前一次输出的第一分片点云数据；所述第一分片点云数据为所述雷达每旋转一个预设角度所采集到的点云数据；所述预设角度小于360度；
从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；所述待拼接点云数据为所述雷达前一次输出的第一分片点云数据中、与所述雷达当前输出的第一分片点云数据相邻的部分点云数据；
将所述待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；
对所述第二分片点云数据进行检测，得到目标检测结果。
本公开实施例，将前一次获取的第一分片点云数据中的部分点云数据与当前获取的第一分片点云数据进行拼接，从而得到拼接处附近的全面的信息，因此，能够克服点云数据分割处漏检的问题，提高检测的准确率。另外，将雷达采集的点云数据分片处理，能够减少雷达单次处理的数据量，从而能够提高检测效率，在应用于自动驾驶***中，能够满足实时性获取检测结果的要求。
一种可选的实施方式中，所述从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据，包括：
确定前一次输出的第一分片点云数据的第一边界；所述第一边界为所述雷达采集所述第一分片点云数据结束时形成的点云边界，且所述第一边界与所述雷达的旋转方向相垂直；
从前一次输出的第一分片点云数据中，筛选与所述第一边界的距离小于第一预设距离的点云数据，并将筛选得到的点云数据作为所述待拼接点云数据。
该实施方式，基于第一边界从前一次输出的第一分片点云数据中，筛选与当前获取的第一分片点云数据相接的点云数据作为待拼接点云数据，能够补全点云分割处的点云数据，从而得到拼接处附近的全面的信息，进而克服了点云数据分割处漏检的问题。
一种可选的实施方式中，所述将所述待拼接点云数据与当前输出的所述第一分片点云数据进行拼接，得到第二分片点云数据，包括：
确定当前输出的第一分片点云数据的第二边界；所述第二边界为所述雷达开始采集所述第一分片点云数据时所对应的边界点云数据，且所述第二边界与所述雷达的旋转方向相垂直；
将所述待拼接点云数据拼接在当前输出的第一分片点云数据中的所述第二边界侧，得到第二分片点云数据。
该实施方式，利用第二边界，实现了准确地将待拼接点云数据拼接在当前输出的第一分片点云数据上，形成一个能够覆盖当前获取的第一分片点云数据的第二分片点云数据，可以解决当前输出的第一分片点云数据分割处漏检的问题。
一种可选的实施方式中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
在对所述第二分片点云数据进行检测，得到目标检测结果之后，所述方法还包括：
从前一次输出的第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组；
将所述第一检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除。
该实施方式，剔除第一检测框组中除最高置信度对应的对象检测框以外的对象检测框，有效降低了错误检测率，提高了检测精度。
一种可选的实施方式中，所述从前一次输出的所述第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组，包括：
从前一次输出的第一分片点云数据对应的所述对象检测框，和所述第二分片点云数据对应的所述对象检测框中，获取至少部分重叠的两个对象检测框；
在所述至少部分重叠的两个对象检测框的重叠区域面积，与所述至少部分重叠的两个对象检测框中任一个所述对象检测框的面积的比值大于预设值的情况下，将所述至少部分重叠的两个对象检测框作为一个所述第一检测框组。
该实施方式，通过两个重叠的对象检测框的重叠区域面积能够将大概率属于同一个对象的对象检测框作为一个第一检测框组，以剔除其中的置信度 低的对象检测框，而将置信度最高的对象检测框作为对应对象最终的对象检测框，避免同一对象存在两个对象检测框的情况，能够有效降低错误检测率，提高检测精度；另外，重叠区域面积较小的两个对象检测框大概率属于不同的对象，无需对此种情况进行对象检测框的处理，进一步提高了检测精度。
一种可选的实施方式中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
所述对所述第二分片点云数据进行检测，得到目标检测结果之后，所述方法还包括：
确定所述第二分片点云数据的第三边界；所述第三边界与所述雷达的旋转方向相垂直；
从所述第二分片点云数据对应的所述对象检测框中筛选与所述第三边界的距离小于第二预设距离的第一目标检测框；
基于预先设置的衰减比例因子、所述第一目标检测框的中心点到所述第三边界的距离、以及所述第二预设距离，对所述第一目标检测框的置信度进行修正。
该实施方式，由于距离边界处较近的对象检测框的置信度实际上较低，但是预测得到的检测框的置信度较高，利用该实施方式能够将距离边界处较近的对象检测框的置信度修正为与实际情况相符的置信度，进而能够得到更为准确的第三边界侧的目标检测结果。
第二方面，本公开实施例还提供一种目标检测装置，包括：
获取部分，被配置为获取雷达当前输出的第一分片点云数据以及所述雷达前一次输出的第一分片点云数据；所述第一分片点云数据为所述雷达每旋转一个预设角度所采集到的点云数据；所述预设角度小于360度；
提取部分，被配置为从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；所述待拼接点云数据为所述雷达前一次输出的第一分片点云数据中、与所述雷达当前输出的第一分片点云数据相邻的部分点云数据；
拼接部分，被配置为将所述待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；
检测部分，被配置为对所述第二分片点云数据进行检测，得到目标检测结果。
一种可选的实施方式中，所述提取部分，被配置为确定前一次输出的第一分片点云数据的第一边界；所述第一边界为所述雷达采集所述第一分片点云数据结束时形成的点云边界，且所述第一边界与所述雷达的旋转方向相垂直；
从前一次输出的第一分片点云数据中，筛选与所述第一边界的距离小于第一预设距离的点云数据，并将筛选得到的点云数据作为所述待拼接点云数据。
一种可选的实施方式中，所述拼接部分，被配置为确定当前输出的第一分片点云数据的第二边界；所述第二边界为所述雷达开始采集所述第一分片点云数据时所对应的边界点云数据，且所述第二边界与所述雷达的旋转方向 相垂直；
将所述待拼接点云数据拼接在当前输出的第一分片点云数据中的所述第二边界侧，得到第二分片点云数据。
一种可选的实施方式中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
所述检测部分，还被配置为在对所述第二分片点云数据进行检测，得到目标检测结果之后，从前一次输出的第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组；
将所述第一检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除。
一种可选的实施方式中，所述检测部分，被配置为从前一次输出的第一分片点云数据对应的所述对象检测框，和所述第二分片点云数据对应的所述对象检测框中，获取至少部分重叠的两个对象检测框；在所述至少部分重叠的两个对象检测框的重叠区域面积，与所述至少部分重叠的两个对象检测框中任一个所述对象检测框的面积的比值大于预设值的情况下，将所述至少部分重叠的两个对象检测框作为一个所述第一检测框组。
一种可选的实施方式中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；所述检测部分，还被配置为在对所述第二分片点云数据进行检测，得到目标检测结果之后，确定所述第二分片点云数据的第三边界；所述第三边界与所述雷达的旋转方向相垂直；从所述第二分片点云数据对应的所述对象检测框中筛选与所述第三边界的距离小于第二预设距离的第一目标检测框；
基于预先设置的衰减比例因子、所述第一目标检测框的中心点到所述第三边界的距离、以及所述第二预设距离，对所述第一目标检测框的置信度进行修正。
第三方面，本公开实施例还提供一种计算机设备，包括：处理器、存储器和总线，所述存储器存储有所述处理器可执行的机器可读指令，当计算机设备运行时，所述处理器与所述存储器之间通过总线通信，所述机器可读指令被所述处理器执行时执行上述第一方面，或第一方面中任一种可能的实施方式中的步骤。
第四方面，本公开实施例还提供一种计算机可读存储介质，该计算机可读存储介质上存储有计算机程序，该计算机程序被处理器运行时执行上述第一方面，或第一方面中任一种可能的实施方式中的步骤。
第五方面，本公开实施例提供一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行如上述第一方面，或第一方面中任一种可能的实施方式中的步骤。
关于上述目标检测装置、计算机设备和存储介质的效果描述参见上述目标检测方法的说明。
本公开实施例提供的目标检测方法、装置、计算机设备和存储介质，通 过获取雷达当前输出的第一分片点云数据以及所述雷达前一次输出的第一分片点云数据；第一分片点云数据为雷达每旋转一个预设角度所采集到的点云数据；预设角度小于360度；从雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；待拼接点云数据为雷达前一次输出的第一分片点云数据中、与雷达当前输出的第一分片点云数据相邻的部分点云数据；将待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；对第二分片点云数据进行检测，得到目标检测结果，与现有技术中在满足实时输出检测结果的情况下所导致的检测结果精度低相比，其通过将前一次获取的第一分片点云数据中的部分点云数据与当前获取的第一分片点云数据进行拼接，从而得到拼接处附近的全面的信息，因此，能够克服点云数据分割处漏检的问题，进一步提高检测的准确率。另外，将雷达采集的点云数据分片处理，能够减少雷达单次处理的数据量，从而能够提高检测效率，在应用于自动驾驶***中，能够满足实时性获取检测结果的要求。
附图说明
为了更清楚地说明本公开实施例的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，此处的附图被并入说明书中并构成本说明书中的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。应当理解，以下附图仅示出了本公开的某些实施例，因此不应被看作是对范围的限定，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他相关的附图。
图1示出了本公开实施例所提供的一种目标检测方法的流程图；
图2示出了本公开实施例所提供的雷达分片输出第一分片点云数据的示意图；
图3示出了本公开实施例所提供的得到第二分片点云数据的流程示意图；
图4示出了本公开实施例所提供的剔除拼接后的目标检测结果中重复的检测结果的流程图；
图5示出了本公开实施例所提供的第一分片点云数据以及第二分片点云数据中对象检测框的分布示意图；
图6示出了本公开实施例所提供的修正第一目标对测框的置信度的流程图；
图7示出了本公开实施例所提供的一种目标检测装置的示意图；
图8示出了本公开实施例所提供的一种计算机设备的结构示意图。
具体实施方式
为使本公开实施例的目的、技术方案和优点更加清楚，下面将结合本公开实施例中附图，对本公开实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本公开一部分实施例，而不是全部的实施例。通常在此处附图中描述和示出的本公开实施例的组件可以以各种不同的配置来布置和设计。因此，以下对在附图中提供的本公开的实施例的详细描述并非旨在限制要求保护的本公开的范围，而是仅仅表示本公开的选定实施例。基 于本公开的实施例，本领域技术人员在没有做出创造性劳动的前提下所获得的所有其他实施例，都属于本公开保护的范围。
另外，本公开实施例中的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的实施例能够以除了在这里图示或描述的内容以外的顺序实施。
在本文中提及的“多个或者若干个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。字符“/”一般表示前后关联对象是一种“或”的关系。
经研究发现，雷达感知算法往往基于参数量巨大的深度神经网络，很难满足自动驾驶***实时性的要求；另外，通过复杂的压缩加速的雷达感知算法还会在一定程度上牺牲神经网络模型精度。
针对以上方案所存在的缺陷，均是发明人在经过实践并仔细研究后得出的结果，因此，上述问题的发现过程以及下文中本公开针对上述问题所提出的解决方案，都属于本公开实施例保护的范围。
应注意到：相似的标号和字母在下面的附图中表示类似项，因此，一旦某一项在一个附图中被定义，则在随后的附图中不需要对其进行进一步定义和解释。
为便于对本实施例进行理解，首先对本公开实施例所公开的一种目标检测方法的应用场景进行介绍，首先介绍雷达在应用场景中的工作过程：
雷达采集的点云数据是通过雷达发射器发射激光，遇到障碍物反射，再由雷达接收器接受激光后，雷达经过运算获得障碍物表面发射点的集合。激光遇到障碍物并返回被接收器接受的最远距离决定了雷达可以检测的最远范围。雷达一般配有多个成对的发射器和接收器，不同的发射器和接收器向不同的角度发射或接受激光。
这里，雷达可以包括机械式激光雷达。机械式激光雷达的发射器还可以一边发射和接受激光，一边沿着固定旋转轴旋转，以获得360度全方位的点云数据信息。
以雷达旋转360度为一个整周期，雷达在获得一个完整周期的点云数据(即雷达采集的一整帧数据)后，可以通过数据端口将整帧数据发送给后续模块(比如神经网络模块)来处理，最终呈现的点云数据是以雷达旋转中心为中心，最远检测范围为半径的一个圆柱体区域内的点云数据的集合，圆柱体的高由不同发射接受器之间的最大角度决定。实际应用中可以采集的点云数据范围小于规则的圆柱体。需要说明的是，由于本公开实施例应用的自动驾驶场景中不考虑雷达正上方(比如，天空)和雷达正下方(比如车顶或者地面等)的圆锥型区域，因此本公开实施例涉及到的雷达的检测范围认为是一个圆柱体。
本发明实施例提供的目标检测方法可以应用于自动驾驶应用场景中，在自动驾驶应用场景中，通常将雷达安装在车顶处，由于处理点云数据的神经 网络模型擅长处理矩形区域内的信息，结合自动驾驶应用场景，需要将雷达采集到的原始点云数据做进一步的裁剪处理，能够得到以雷达为中心的一个长方体，该长方体与地面平行，长方体的长边沿着车头方向，宽边垂直于车头方向，高边垂直于地面。雷达采集点云数据，并输出点云数据所在长方体区域内的障碍物种类和位置信息。使用神经网络模型检测长方体区域的点云数据，并输出长方体区域内的检测结果，上述长方体即为下述对象检测框。
综上，可以知道雷达旋转360度采集的点云数据，通过神经网络模型进行点云数据处理，最终输出360度检测结果。
为便于对本实施例进行理解，首先对本公开实施例所公开的一种目标检测方法进行详细介绍，本公开实施例所提供的目标检测方法的执行主体一般为具有一定计算能力的计算机设备，该计算机设备例如包括：终端设备或服务器或其它处理设备，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、终端、蜂窝电话、无绳电话、个人数字助理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等。在一些可能的实现方式中，该目标检测方法可以通过处理器调用存储器中存储的计算机可读指令的方式来实现。
下面以执行主体为计算机设备为例对本公开实施例提供的目标检测方法加以说明。
基于上述自动驾驶应用场景，参见图1所示，为本公开实施例提供的一种目标检测方法的流程图，该方法包括步骤S101至S104，其中：
S101：获取雷达当前输出的第一分片点云数据以及雷达前一次输出的第一分片点云数据。
其中，第一分片点云数据为雷达每旋转一个预设角度所采集到的点云数据。这里，预设角度小于360度。
本步骤中，预设角度可以根据神经网络模型的运算量，以及应用场景中测试对象的复杂度，自定义所需要的角度值，该角度一般小于雷达旋转一个周期的全角度，在此不进行限定。
示例性的，在自动驾驶应用场景中，通过机械式激光雷达采集的点云数据具有时序信息，即通过360度旋转激光发射和接收器，可以连续采集点云数据。可以以机械式激光雷达旋转一个预设角度为一个子周期，采集点云数据，以机械式激光雷达旋转360度为整周期，这里子周期是通过将机械式激光雷达旋转360度整周期均分为若干份来获得的。参见图2所示，其为雷达分片输出第一分片点云数据的示意图，其中，以每90度为一个子周期，输出一片第一分片点云数据。
S102：从雷达前一次输出的第一分片点云数据中，提取待拼接点云数据。
本步骤中，待拼接点云数据为雷达前一次输出的第一分片点云数据中、与雷达当前输出的第一分片点云数据相邻的部分点云数据。
在实施的过程中，首先确定前一次输出的第一分片点云数据的第一边界；其中，第一边界可以为雷达采集第一分片点云数据结束时形成的点云边界，且第一边界与雷达的旋转方向相垂直；从前一次输出的第一分片点云数据中， 筛选与第一边界的距离小于第一预设距离的点云数据，并将筛选得到的点云数据作为待拼接点云数据，这样能够补全点云分割处的点云数据，从而克服点云数据分割处的漏检的问题。
其中，第一预设距离在不同的应用场景中可以根据经验值自定义，在此不进行限定。示例性的，在自动驾驶应用场景中，第一预设距离可以根据检测对象的尺寸数据定义，比如车长6米，则第一预设距离可以选取大于6米的数据，该数据可以根据经验值和检测对象实际尺寸数据综合考虑的权衡结果进行定义，在此不限定。
参见图3所示，其为得到第二分片点云数据的流程示意图。其中，雷达前一次输出的第一分片点云数据31、雷达当前输出的第一分片点云数据32。
针对第一分片点云数据31，根据雷达的旋转方向，如图3所示，即顺时针旋转，确定出第一分片点云数据31的第一边界311，从第一分片点云数据31中，筛选与第一边界311的距离小于第一预设距离的点云数据，并将筛选的到的点云数据作为待拼接点云数据312。
S103：将待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据。
在实施的过程中，确定当前输出的第一分片点云数据的第二边界；第二边界为雷达开始采集第一分片点云数据时所对应的边界点云数据，且第二边界与雷达的旋转方向相垂直；将待拼接点云数据拼接在当前输出的第一分片点云数据中的第二边界侧，得到第二分片点云数据。上述利用第二边界，能够实现准确地将待拼接点云数据拼接在当前输出的第一分片点云数据上，形成一个能够覆盖当前获取的第一分片点云数据的第二分片点云数据，可以解决当前输出的第一分片点云数据分割处漏检的问题。
这里，基于图3所示，针对第一分片点云数据32，根据雷达的旋转方向，确定出第一分片点云数据32的第二边界321，这里，待拼接点云数据312存在与第一分片点云数据31相同的边界，即第一边界311，将待拼接点云数据312的第一边界311侧拼接在第一分片点云数据32中的第二边界321侧，得到第二分片点云数据33。
S104：对第二分片点云数据进行检测，得到目标检测结果。
在实施的过程中，对第二分片点云数据进行检测，可以使用神经网络模型进行检测，得到神经网络模型输出的目标检测结果。
雷达输出采集到的第一个子周期内的第一分片点云数据，也可以使用神经网络模型进行检测，得到神经网络模型输出的目标检测结果。
这里，针对得到的目标检测结果。一种实施方式可以为，可以将神经网络模型输出的、雷达一个整周期的目标检测结果拼接，得到全角度的目标检测结果。另一种实施方式为，将神经网络模型输出的前一次目标检测结果与本次目标检测结果拼接，得到部分角度的目标检测结果。上述两种方式都可以解决当前输出的第一分片点云数据拼接处检测结果漏检的问题。
针对上述拼接后的目标检测结果进行处理，可以参见图4所示，其为剔除拼接后的目标检测结果中重复的检测结果的流程图。其中：
S401：从前一次输出的第一分片点云数据对应的对象检测框，和第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组。
本步骤中，前一次输出的第一分片点云数据对应的对象检测框，可以有以下多种确定方式：
方式1、在前一次输出的第一分片点云数据为雷达输出的第一片点云数据的情况下，使用神经网络模型检测上述前一次输出的第一分片点云数据，得到前一次输出的第一分片点云数据对应的对象检测框。
方式2、在前一次输出的第一分片点云数据为雷达输出的第一片点云数据的情况下，使用神经网络模型检测前一次输出的第二分片点云数据(前一次输出的第一分片点云数据为前一次输出的第二分片点云数据中的一部分)，将得到的对象检测框作为前一次输出的第一分片点云数据对应的对象检测框。
在实施的过程中，可以从前一次输出的第一分片点云数据对应的对象检测框，和第二分片点云数据对应的对象检测框中，获取至少部分重叠的两个对象检测框；在至少部分重叠的两个对象检测框的重叠区域面积，与至少部分重叠的两个对象检测框中任一个对象检测框的面积的比值大于预设值的情况下，将至少部分重叠的两个对象检测框作为一个第一检测框组。这里，重叠区域面积较小的两个对象检测框大概率属于不同的对象，无需对此种情况进行对象检测框的处理。
其中，预设值在不同的应用场景中可以根据经验值自定义，在此不进行限定。
示例性的，可以参见图5所示，其为第一分片点云数据以及第二分片点云数据中对象检测框的分布示意图。以上述方式1为例，确定出第一分片点云数据50中的对象检测框的分布位置，其中，在神经网络模型识别出待拼接点云数据51内存在对象检测框511、对象检测框512和对象检测框523为不完全检测框的情况下，神经网络模型可以预测出其剩余部分检测框，即对象检测框511中的虚线部分、对象检测框512中的虚线部分和对象检测框523中的虚线部分。可以从前一次输出的第一分片点云数据50中的对象检测框，和第二分片点云数据52对应的对象检测框中，获取至少部分重叠的两个对象检测框，其中对象检测框511与对象检测框521、对象检测框512与对象检测框522、对象检测框513与对象检测框523、对象检测框514与对象检测框524、对象检测框511与对象检测框514为部分重叠的两个对象检测框，根据预设值大小，可以确定出对象检测框511与对象检测框521、对象检测框512与对象检测框522、对象检测框513与对象检测框523、对象检测框514与对象检测框524满足两个对象检测框的重叠区域面积，与至少部分重叠的两个对象检测框中任一个对象检测框的面积的比值大于预设值的条件，上述四组对象检测框都可以作为第一检测框组。对象检测框511与对象检测框514重叠区域面积与对象检测框511或对象检测框514任一个对象检测框的面积比值小于或等于预设值，所以对象检测框511与对象检测框514不能作为第一检测框组。
S402：将第一检测框组中除最高置信度对应的对象检测框以外的对象检 测框剔除。
这里，目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度。
根据非极大值抑制算法的原理，基于步骤S401中得到的第一检测框组，以及第一检测框组中每一对象检测框的置信度，确定第一对象检测框组中置信度最高的对象检测框，将该置信度最高的对象检测框保留，剔除该第一检测框组中剩余对象检测框。这里，针对第一检测框组，剔除其中的置信度低的对象检测框，而将置信度最高的对象检测框作为对应对象最终的对象检测框，避免同一对象存在两个对象检测框的情况，能够有效降低错误检测率，提高检测精度。
针对拼接后的目标检测结果进行处理，在剔除拼接后的目标检测结果中重复的检测结果之前，还可以对第一目标检测框的置信度进行修正。可以参见图6所示，其为修正第一目标对测框的置信度的流程图，其中：
S601：确定第二分片点云数据的第三边界；第三边界与雷达的旋转方向相垂直。
本步骤中，第二分片点云数据的第三边界可以为图3中的边界331和边界332。
S602：从第二分片点云数据对应的对象检测框中筛选与第三边界的距离小于第二预设距离的第一目标检测框。
本步骤中，第二预设距离在不同的应用场景中可以根据经验值自定义，在此不进行限定。
S603：基于预先设置的衰减比例因子、第一目标检测框的中心点到第三边界的距离、以及第二预设距离，对第一目标检测框的置信度进行修正。
本步骤中，预先设置的衰减比例因子可以取0至1范围内的数值。
在实施的过程中，可以根据公式(1)示出的线性衰减函数确定：
new_score＝old_score×(1-scale_factor×max(0,(D-d)/D))公式(1)；
对第一目标检测框的置信度进行修正，其中，new_score表示第一目标检测框修正后的置信度，old_score表示神经网络模型输出的第一目标检测框的置信度，scale_factor表示衰减比例因子，D表示第二预设距离，d表示第一目标检测框的中心到第三边界的距离，max(0,(D-d)/D)表示0和(D-d)/D之间取最大值。
在一种可能的实施方式中，在获取雷达当前输出的第一分片点云数据以及雷达前一次输出的第一分片点云数据之后，还可以对检测第一分片点云数据得到的点云检测结果(即第一点云检测结果和第二点云检测结果)直接进行去重复处理。在实施的过程中，对雷达当前输出的第一分片点云数据进行检测，得到第一点云检测结果；对雷达前一次输出的第一分片点云数据进行检测，得到第二点云检测结果；从第一点云检测结果对应的对象检测框，和第二点云检测结果对应的对象检测框中，获取至少部分重叠的第二检测框组；将第二检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除。 上述实施过程的进一步细节说明可以参照步骤S401至S402。
这里，在将第二检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除之前，还需要对对象检测框的置信度进行修正。在实施的过程中，确定第一分片点云数据的第四边界；第四边界与雷达的旋转方向相垂直；从第一分片点云数据对应的对象检测框中筛选与第四边界的距离小于第三预设距离的第二目标检测框；基于预先设置的衰减比例因子、第二目标检测框的中心点到第四边界点的距离、以及第三预设距离，对第二目标检测框的置信度进行修正。上述实施过程的进一步细节说明可以参照步骤S601至S603。
上述步骤S101至步骤S104，利用雷达分片输出的第一分片点云数据，将从前一次输出的第一分片点云数据中获得的待拼接点云数据拼接到当前输出的第一分片点云数据中，得到新的分片点云数据，即第二分片点云数据，得到的第二分片点云数据包括了当前输出的第一分片点云数据拼接处的全部点云数据，因此，通过将前一次获取的第一分片点云数据中的部分点云数据与当前获取的第一分片点云数据进行拼接，能够得到点云数据分割处附近的全局信息，因此，能够克服点云数据分割处漏检的问题，进一步提高检测的准确率。另外，将雷达采集的点云数据分片处理，能够减少雷达单次处理的数据量，从而能够提高检测效率，在应用于自动驾驶***中，能够满足实时获取检测结果的要求。
本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的撰写顺序并不意味着严格的执行顺序而对实施过程构成任何限定，各步骤的执行顺序应当以其功能和可能的内在逻辑确定。
基于同一发明构思，本公开实施例中还提供了与目标检测方法对应的目标检测装置，由于本公开实施例中的目标检测装置解决问题的原理与本公开实施例上述目标检测方法相似，因此装置的实施可以参见方法的实施。
参照图7所示，为本公开实施例提供的一种目标检测装置的示意图，所述装置包括：获取部分701、提取部分702、拼接部分703、检测部分704；其中，
获取部分701，被配置为获取雷达当前输出的第一分片点云数据以及所述雷达前一次输出的第一分片点云数据；所述第一分片点云数据为所述雷达每旋转一个预设角度所采集到的点云数据；所述预设角度小于360度；
提取部分702，被配置为从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；所述待拼接点云数据为所述雷达前一次输出的第一分片点云数据中、与所述雷达当前输出的第一分片点云数据相邻的部分点云数据；
拼接部分703，被配置为将所述待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；
检测部分704，被配置为对所述第二分片点云数据进行检测，得到目标检测结果。
一种可选的实施方式中，所述提取部分702，被配置为确定前一次输出的第一分片点云数据的第一边界；所述第一边界为所述雷达采集所述第一分片 点云数据结束时形成的点云边界，且所述第一边界与所述雷达的旋转方向相垂直；
从前一次输出的第一分片点云数据中，筛选与所述第一边界的距离小于第一预设距离的点云数据，并将筛选得到的点云数据作为所述待拼接点云数据。
一种可选的实施方式中，所述拼接部分703，被配置为确定当前输出的第一分片点云数据的第二边界；所述第二边界为所述雷达开始采集所述第一分片点云数据时所对应的边界点云数据，且所述第二边界与所述雷达的旋转方向相垂直；
将所述待拼接点云数据拼接在当前输出的第一分片点云数据中的所述第二边界侧，得到第二分片点云数据。
一种可选的实施方式中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
所述检测部分704，还被配置为在对所述第二分片点云数据进行检测，得到目标检测结果之后，从前一次输出的第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组；
将所述第一检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除。
一种可选的实施方式中，所述检测部分704，被配置为从前一次输出的第一分片点云数据对应的所述对象检测框，和所述第二分片点云数据对应的所述对象检测框中，获取至少部分重叠的两个对象检测框；在所述至少部分重叠的两个对象检测框的重叠区域面积，与所述至少部分重叠的两个对象检测框中任一个所述对象检测框的面积的比值大于预设值的情况下，将所述至少部分重叠的两个对象检测框作为一个所述第一检测框组。
一种可选的实施方式中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；所述检测部分704，还被配置为在对所述第二分片点云数据进行检测，得到目标检测结果之后，确定所述第二分片点云数据的第三边界；所述第三边界与所述雷达的旋转方向相垂直；从所述第二分片点云数据对应的所述对象检测框中筛选与所述第三边界的距离小于第二预设距离的第一目标检测框；
基于预先设置的衰减比例因子、所述第一目标检测框的中心点到所述第三边界的距离、以及所述第二预设距离，对所述第一目标检测框的置信度进行修正。
关于目标检测装置中的各部分的处理流程、以及各部分之间的交互流程的描述可以参照上述目标检测方法实施例中的相关说明，这里不再详述。
基于同一技术构思，本申请实施例还提供了一种计算机设备。参照图8所示，为本申请实施例提供的一种计算机设备的结构示意图，包括：
处理器81、存储器82和总线83。其中，存储器82存储有处理器81可执行的机器可读指令，处理器81被配置为执行存储器82中存储的机器可读 指令，所述机器可读指令被处理器81执行时，处理器81执行下述步骤：
获取雷达当前输出的第一分片点云数据以及雷达前一次输出的第一分片点云数据；所述第一分片点云数据为所述雷达每旋转一个预设角度所采集到的点云数据；所述预设角度小于360度；
从雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；所述待拼接点云数据为所述雷达前一次输出的第一分片点云数据中、与所述雷达当前输出的第一分片点云数据相邻的部分点云数据；
将待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；
对第二分片点云数据进行检测，得到目标检测结果。
上述存储器82包括内存821和外部存储器822；这里的内存821也称内存储器，被配置为暂时存放处理器81中的运算数据，以及与硬盘等外部存储器822交换的数据，处理器81通过内存821与外部存储器822进行数据交换，当计算机设备运行时，处理器81与存储器82之间通过总线83通信，使得处理器81在执行上述方法实施例中所提及的执行指令。
本公开实施例还提供一种计算机可读存储介质，该计算机可读存储介质上存储有计算机程序，该计算机程序被处理器运行时执行上述方法实施例中所述的目标检测方法的步骤。其中，该存储介质可以是易失性或非易失的计算机可读取存储介质。
本公开实施例所还提供一种计算机程序产品，该计算机程序产品承载有程序代码，所述程序代码包括的指令可用于执行上述方法实施例中所述的目标检测方法的步骤，可参见上述方法实施例。
其中，上述计算机程序产品可以通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品可以体现为计算机存储介质，在另一个可选实施例中，计算机程序产品可以体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的***和装置的工作过程，可以参考前述方法实施例中的对应过程。在本公开所提供的几个实施例中，应该理解到，所揭露的***、装置和方法，可以通过其它的方式实现。以上所描述的装置实施例仅仅是示意性的，例如，所述部分的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，又例如，多个部分或组件可以结合或者可以集成到另一个***，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些通信接口，装置或部分的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本公开各个实施例中的各功能部分可以集成在一个处理部分中， 也可以是各个部分单独物理存在，也可以两个或两个以上部分集成在一个部分中。
所述功能如果以软件功能部分的形式实现并作为独立的产品销售或使用时，可以存储在一个处理器可执行的非易失的计算机可读取存储介质中。基于这样的理解，本公开实施例的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本公开各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
最后应说明的是：以上所述实施例，仅为本公开的实施方式，用以说明本公开实施例的技术方案，而非对其限制，本公开的保护范围并不局限于此，尽管参照前述实施例对本公开进行了详细的说明，本领域的普通技术人员应当理解：任何熟悉本技术领域的技术人员在本公开实施例揭露的技术范围内，其依然可以对前述实施例所记载的技术方案进行修改或可轻易想到变化，或者对其中部分技术特征进行等同替换；而这些修改、变化或者替换，并不使相应技术方案的本质脱离本公开实施例技术方案的精神和范围，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应所述以权利要求的保护范围为准。
工业实用性
本公开提供了一种目标检测的方法、装置、计算机设备和存储介质，其中，该方法包括：获取雷达当前输出的第一分片点云数据以及雷达前一次输出的第一分片点云数据；第一分片点云数据为雷达每旋转一个预设角度所采集到的点云数据；预设角度小于360度；从雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；待拼接点云数据为雷达前一次输出的第一分片点云数据中、与雷达当前输出的第一分片点云数据相邻的部分点云数据；将待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；对第二分片点云数据进行检测，得到目标检测结果。与现有技术中在满足实时输出检测结果的情况下所导致的检测结果精度低相比，其通过将前一次获取的第一分片点云数据中的部分点云数据与当前获取的第一分片点云数据进行拼接，从而得到拼接处附近的全面的信息，因此，能够克服点云数据分割处漏检的问题，进一步提高检测的准确率。另外，将雷达采集的点云数据分片处理，能够减少雷达单次处理的数据量，从而能够提高检测效率，在应用于自动驾驶***中，能够满足实时性获取检测结果的要求。

Claims (11)

  1. 一种目标检测方法，包括：
    获取雷达当前输出的第一分片点云数据以及所述雷达前一次输出的第一分片点云数据；所述第一分片点云数据为所述雷达每旋转一个预设角度所采集到的点云数据；所述预设角度小于360度；
    从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；所述待拼接点云数据为所述雷达前一次输出的第一分片点云数据中、与所述雷达当前输出的第一分片点云数据相邻的部分点云数据；
    将所述待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；
    对所述第二分片点云数据进行检测，得到目标检测结果。
  2. 根据权利要求1所述的方法，其中，所述从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据，包括：
    确定前一次输出的第一分片点云数据的第一边界；所述第一边界为所述雷达采集所述第一分片点云数据结束时形成的点云边界，且所述第一边界与所述雷达的旋转方向相垂直；
    从前一次输出的第一分片点云数据中，筛选与所述第一边界的距离小于第一预设距离的点云数据，并将筛选得到的点云数据作为所述待拼接点云数据。
  3. 根据权利要求1或2所述的方法，其中，所述将所述待拼接点云数据与当前输出的所述第一分片点云数据进行拼接，得到第二分片点云数据，包括：
    确定当前输出的第一分片点云数据的第二边界；所述第二边界为所述雷达开始采集所述第一分片点云数据时所对应的边界点云数据，且所述第二边界与所述雷达的旋转方向相垂直；
    将所述待拼接点云数据拼接在当前输出的第一分片点云数据中的所述第二边界侧，得到第二分片点云数据。
  4. 根据权利要求1至3任一项所述的方法，其中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
    在对所述第二分片点云数据进行检测，得到目标检测结果之后，所述方法还包括：
    从前一次输出的第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组；
    将所述第一检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除。
  5. 根据权利要求4所述的方法，其中，所述从前一次输出的第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组，包括：
    从前一次输出的第一分片点云数据对应的所述对象检测框，和所述第二分片点云数据对应的所述对象检测框中，获取至少部分重叠的两个对象检测 框；
    在所述至少部分重叠的两个对象检测框的重叠区域面积，与所述至少部分重叠的两个对象检测框中任一个所述对象检测框的面积的比值大于预设值的情况下，将所述至少部分重叠的两个对象检测框作为一个所述第一检测框组。
  6. 根据权利要求1至5任一项所述的方法，其中，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
    所述对所述第二分片点云数据进行检测，得到目标检测结果之后，所述方法还包括：
    确定所述第二分片点云数据的第三边界；所述第三边界与所述雷达的旋转方向相垂直；
    从所述第二分片点云数据对应的所述对象检测框中筛选与所述第三边界的距离小于第二预设距离的第一目标检测框；
    基于预先设置的衰减比例因子、所述第一目标检测框的中心点到所述第三边界的距离、以及所述第二预设距离，对所述第一目标检测框的置信度进行修正。
  7. 一种目标检测装置，包括：
    获取部分，被配置为获取雷达当前输出的第一分片点云数据以及所述雷达前一次输出的第一分片点云数据；所述第一分片点云数据为所述雷达每旋转一个预设角度所采集到的点云数据；所述预设角度小于360度；
    提取部分，被配置为从所述雷达前一次输出的第一分片点云数据中，提取待拼接点云数据；所述待拼接点云数据为所述雷达前一次输出的第一分片点云数据中、与所述雷达当前输出的第一分片点云数据相邻的部分点云数据；
    拼接部分，被配置为将所述待拼接点云数据与当前输出的第一分片点云数据进行拼接，得到第二分片点云数据；
    检测部分，被配置为对所述第二分片点云数据进行检测，得到目标检测结果。
  8. 根据权利要求7所述的装置，所述目标检测结果包括至少一个对象检测框以及每个对象检测框的置信度；
    所述检测部分，还被配置为在对所述第二分片点云数据进行检测，得到目标检测结果之后，从前一次输出的第一分片点云数据对应的对象检测框，和所述第二分片点云数据对应的对象检测框中，获取至少部分重叠的第一检测框组；
    将所述第一检测框组中除最高置信度对应的对象检测框以外的对象检测框剔除。
  9. 一种计算机设备，包括：处理器、存储器和总线，所述存储器存储有所述处理器可执行的机器可读指令，当计算机设备运行时，所述处理器与所述存储器之间通过总线通信，所述机器可读指令被所述处理器执行时执行如权利要求1至6任一所述的目标检测方法的步骤。
  10. 一种计算机可读存储介质，所述计算机可读存储介质上存储有计算 机程序，所述计算机程序被处理器运行时执行如权利要求1至6任一项所述的目标检测方法的步骤。
  11. 一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现权利要求1至6任一项所述的目标检测方法的步骤。
PCT/CN2021/103186 2021-03-29 2021-06-29 一种目标检测方法、装置、计算机设备和存储介质 WO2022205644A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110332125.4 2021-03-29
CN202110332125.4A CN112950622B (zh) 2021-03-29 2021-03-29 一种目标检测方法、装置、计算机设备和存储介质

Publications (1)

Publication Number Publication Date
WO2022205644A1 true WO2022205644A1 (zh) 2022-10-06

Family

ID=76227126

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/103186 WO2022205644A1 (zh) 2021-03-29 2021-06-29 一种目标检测方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN112950622B (zh)
WO (1) WO2022205644A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116148823A (zh) * 2023-04-12 2023-05-23 北京集度科技有限公司 一种外参数的标定方法、装置、车辆及计算机程序产品

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112950622B (zh) * 2021-03-29 2023-04-18 上海商汤临港智能科技有限公司 一种目标检测方法、装置、计算机设备和存储介质
CN113469042A (zh) * 2021-06-30 2021-10-01 上海商汤临港智能科技有限公司 真值数据确定、神经网络训练、行驶控制方法及装置
CN113986504A (zh) * 2021-10-29 2022-01-28 上海商汤临港智能科技有限公司 一种点云数据的处理方法、装置、电子设备及存储介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TW201643814A (zh) * 2015-01-09 2016-12-16 鴻海精密工業股份有限公司 點雲拼接方法及系統
CN109919989A (zh) * 2019-03-27 2019-06-21 广东工业大学 一种雷达电力巡检的点云数据配准方法、装置及设备
WO2019230123A1 (ja) * 2018-05-30 2019-12-05 株式会社Ihi 検知装置及び検知システム
CN110544201A (zh) * 2019-08-30 2019-12-06 武汉中海庭数据技术有限公司 一种车载激光扫描点云的大范围拼接方法及装置
CN111724558A (zh) * 2019-03-21 2020-09-29 杭州海康威视数字技术股份有限公司 一种监控方法、装置及入侵报警***
CN112163643A (zh) * 2020-10-30 2021-01-01 上海商汤临港智能科技有限公司 样本生成、神经网络的训练、数据处理方法及装置
CN112200129A (zh) * 2020-10-28 2021-01-08 中国人民解放军陆军航空兵学院陆军航空兵研究所 一种基于深度学习的三维目标检测方法、装置及终端设备
CN112950622A (zh) * 2021-03-29 2021-06-11 上海商汤临港智能科技有限公司 一种目标检测方法、装置、计算机设备和存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109188457B (zh) * 2018-09-07 2021-06-11 百度在线网络技术（北京）有限公司 物体检测框的生成方法、装置、设备、存储介质及车辆
CN112017202B (zh) * 2019-05-28 2024-06-14 杭州海康威视数字技术股份有限公司 点云标注方法、装置及***
CN110517193B (zh) * 2019-06-28 2022-04-12 西安理工大学 一种海底声呐点云数据处理方法
CN110992468B (zh) * 2019-11-28 2020-10-30 贝壳找房(北京)科技有限公司 基于点云数据的建模方法、装置以及设备、存储介质

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TW201643814A (zh) * 2015-01-09 2016-12-16 鴻海精密工業股份有限公司 點雲拼接方法及系統
WO2019230123A1 (ja) * 2018-05-30 2019-12-05 株式会社Ihi 検知装置及び検知システム
CN111724558A (zh) * 2019-03-21 2020-09-29 杭州海康威视数字技术股份有限公司 一种监控方法、装置及入侵报警***
CN109919989A (zh) * 2019-03-27 2019-06-21 广东工业大学 一种雷达电力巡检的点云数据配准方法、装置及设备
CN110544201A (zh) * 2019-08-30 2019-12-06 武汉中海庭数据技术有限公司 一种车载激光扫描点云的大范围拼接方法及装置
CN112200129A (zh) * 2020-10-28 2021-01-08 中国人民解放军陆军航空兵学院陆军航空兵研究所 一种基于深度学习的三维目标检测方法、装置及终端设备
CN112163643A (zh) * 2020-10-30 2021-01-01 上海商汤临港智能科技有限公司 样本生成、神经网络的训练、数据处理方法及装置
CN112950622A (zh) * 2021-03-29 2021-06-11 上海商汤临港智能科技有限公司 一种目标检测方法、装置、计算机设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116148823A (zh) * 2023-04-12 2023-05-23 北京集度科技有限公司 一种外参数的标定方法、装置、车辆及计算机程序产品
CN116148823B (zh) * 2023-04-12 2023-09-19 北京集度科技有限公司 一种外参数的标定方法、装置、车辆及计算机程序产品

Also Published As

Publication number Publication date
CN112950622B (zh) 2023-04-18
CN112950622A (zh) 2021-06-11

Similar Documents

Publication Publication Date Title
WO2022205644A1 (zh) 一种目标检测方法、装置、计算机设备和存储介质
CN109886997B (zh) 基于目标检测的识别框确定方法、装置及终端设备
CN107092829B (zh) 一种基于图像匹配的恶意代码检测方法
CN109766744B (zh) 一种松材线虫病树识别与定位方法及***
US20190362157A1 (en) Keyframe-based object scanning and tracking
CN109740549B (zh) 基于半监督cnn的sar图像目标检测***和方法
US10803174B2 (en) Bit-level data generation and artificial intelligence techniques and architectures for data protection
CN106557770B (zh) 通过比较贝塞尔曲线来标识图像中的形状
CN109815823B (zh) 数据处理方法及相关产品
WO2023155387A1 (zh) 多传感器目标检测方法、装置、电子设备以及存储介质
CN112784724A (zh) 一种车辆变道检测方法、装置、设备及存储介质
CN111563398A (zh) 用于确定目标物的信息的方法和装置
US20180268247A1 (en) System and method for detecting change using ontology based saliency
CN110751354B (zh) 一种异常用户的检测方法和装置
CN105302715B (zh) 应用程序用户界面的获取方法和装置
CN111444833B (zh) 果实测产方法、装置、计算机设备及存储介质
CN112395924B (zh) 一种遥感监测方法及装置
CN109785312B (zh) 一种图像模糊检测方法、***及电子设备
CN114708420A (zh) 基于局部方差和后验概率分类器的视觉定位方法及装置
CN110942008B (zh) 一种基于深度学习的面单信息定位方法和***
US20120070033A1 (en) Methods for object-based identification, sorting and ranking of target detections and apparatuses thereof
CN114399432A (zh) 目标识别方法、装置、设备、介质及产品
CN109547276B (zh) 一种问题定位方法、终端及存储介质
CN113344064A (zh) 事件处理方法和装置
TW202215367A (zh) 影像辨識方法及影像辨識系統

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21934293

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE